solanabr/ai-kit / .claude/skills
solana-ai-kit
FlaggedRouting hub for Solana development. Maps a task to the one reference to read first in the ext/ skill submodules or the kit's local skills.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 47.6 KB |
| address-lookup-tables.md | markdown | 5.7 KB |
| backend-async.md | markdown | 6.0 KB |
| csharp-unity-conventions.md | markdown | 6.9 KB |
| deployment.md | markdown | 15.3 KB |
| keeper-ops.md | markdown | 5.7 KB |
| monitoring.md | markdown | 4.2 KB |
| skill-registry.json | config | 145.7 KB |
skill-registry.json
View on GitHub{
"version": "1.2",
"updated": "2026-10-22",
"tiers": {
"core": "Pinned by the kit as a git submodule under .claude/skills/ext/ and installed by default.",
"extension": "Pinned by the kit the same way and installed on demand: bash .claude/bin/skills.sh add <id>. An entry with a skills list is not a submodule: skills.sh fetches those folders from its source into .claude/skills/<name>/.",
"core-plugin": "A Claude Code plugin the kit pins by commit in its own stbr marketplace (.claude-plugin/marketplace.json) and installs with the kit: as the kit plugin's dependency, or by install.sh in a full install. skills.sh lists it but does not manage it; /doctor reports whether it is installed.",
"add-on": "Entries without a tier are opt-in tools the kit does not pin. Install one only when the user asks."
},
"commit": "Every core, extension and core-plugin entry records the commit it is pinned at — for a core plugin, the sha the kit's marketplace pins: for a submodule, the gitlink validate.sh checks the entry against; for a skills-list pack, the commit skills.sh fetches and asserts against FETCH_HEAD. bash .claude/bin/skills.sh pins rechecks them, --write rewrites them from the gitlinks.",
"triggers": "Display text: the INSTALL WHEN column of skills.sh list, for people and agents choosing a pack. Nothing matches on it.",
"scope": "Skill packs, plugins, template repos and catalogs — things an installer can put on disk. MCP servers are not listed here and never were installable from here: skills.sh writes no .mcp.json and runs no claude mcp add. A default server lives in .mcp.json; an opt-in one is a line in .claude/commands/setup-mcp.md that the user runs.",
"entries": [
{
"id": "safe-ai-skill",
"name": "safe-ai-skill",
"type": "plugin",
"tier": "core-plugin",
"domain": "security",
"commit": "8920686ba919faa9804c6d41297bfa8bcdaed527",
"description": "The kit's security plugin: Claude Code hooks that gate mainnet, authority and value-moving actions in Bash and in MCP tools, hold transfers to spend caps, block secret reads, and pin installed skills and MCPs at session start. Installed with the kit, as the kit plugin's dependency or by install.sh; skills.sh does not manage it.",
"triggers": ["always on", "mainnet gate", "spend caps", "secret reads", "skill and MCP vetting"],
"source": "https://github.com/solanabr/safe-ai-skill",
"install": { "method": "plugin-cli", "command": "claude plugin install safe-ai-skill@stbr --scope project", "env": [] },
"maintainer": "solanabr",
"default_installed": true,
"safety": "runs on every Bash, Read, Grep, Glob, Write and Edit, WebFetch, Skill and MCP call, every prompt and every session start, through a prebuilt engine binary its shim picks by platform (macOS and Linux on x64 and arm64, 64-bit Windows under Git Bash). Where no build matches it fails open with a notice, and Solana CLI calls still ask. Its Bash, Skill and permission-request hooks block when the engine fails, so a crashing engine blocks those calls. Its session-start sweep denies a skill pack in place, never moving it, when the pack's files changed since it pinned them or a check finds a BLOCK, until the user runs `safe-ai-skill verify approve <pack>` in a terminal. Its known holes are listed in its SECURITY.md.",
"tags": ["security", "hooks", "mainnet", "spend-caps", "supply-chain", "core"]
},
{
"id": "solana-dev",
"name": "Solana Dev Skill",
"type": "skill",
"tier": "core",
"domain": "solana-core",
"path": ".claude/skills/ext/solana-dev",
"commit": "674da5a4f2e071a61b89628871a462c20ad1928f",
"description": "Solana Foundation development skill: Anchor 1.x and Pinocchio programs, @solana/kit clients, LiteSVM, Mollusk and Surfpool testing, security, Token-2022, payments. The hub's entry point for general Solana work; most agents and commands read it.",
"triggers": ["Solana programs", "Anchor", "Pinocchio", "@solana/kit", "program tests", "security review"],
"source": "https://github.com/solana-foundation/solana-dev-skill",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add solana-dev", "env": [] },
"license": "MIT",
"maintainer": "solana-foundation",
"signal": { "stars": 559, "last_commit": "2026-09-21", "reputability": "official" },
"default_installed": true,
"safety": "three executables and nothing else from the mechanical set. install.sh is for standalone installs and no kit installer runs it; it defaults to a user-scope install and rm -rf's then replaces both $HOME/.agents/skills/solana-dev and $HOME/.claude/skills/solana-dev — the top-level path Claude Code auto-discovers — prompting first on a terminal but overwriting silently when piped, which is the curl route its own README gives. --project writes the same two paths under the cwd and --link symlinks the source instead. The other two, scripts/validate-skill.py and scripts/validate-skills.sh, are the repo's own frontmatter linters. No hooks.json, no .claude/settings.json, no .mcp.json, no symlinks, no package.json and so no install lifecycle hook, no .gitmodules of its own, no nested .claude/skills/, no root CLAUDE.md or AGENTS.md, no bidi or invisible Unicode. No credential read: the keypair, id.json and ~/.config/solana mentions are prose about the reader's own wallet. Four reference files tell you to run curl -sL https://run.surfpool.run/ | bash — a pipe-to-shell it asks you to run, not one it runs.",
"tags": ["solana", "anchor", "pinocchio", "kit", "testing", "security", "official"]
},
{
"id": "auditor-skill",
"name": "Auditor Skill",
"type": "skill",
"tier": "core",
"domain": "security",
"path": ".claude/skills/ext/auditor-skill",
"commit": "c6a4a17297f8994dfb1bb9f2e9e0d19acaab421a",
"vendored": { "vendor/trailofbits": "cfe5d7b1619e47fb5b38b7e2561dad7e5f1e89af" },
"description": "The kit's security pack: 20 checklists over 1,424 verification items and 138 known attack vectors, covering Solana programs (account validation, access control, arithmetic, CPI/PDA, state machines, economics, opsec) and the code around them (TypeScript, Rust services, Python, backend, frontend, supply chain, secrets, deployment, logging, privacy, AI-agent surface). Scope-gated: it loads only the checklists a repo needs. Read by /audit-solana, /audit-infra, /diff-review and the Anchor and Pinocchio agents. It supersedes the trailofbits, ghostsecurity, defending-code and safe-solana-builder packs the kit used to pin; Trail of Bits' tooling is still reachable through its own vendor/trailofbits submodule, which it invokes for execution rather than copying.",
"triggers": ["security audit", "program audit", "vulnerability review", "audit report", "differential review of a PR"],
"source": "https://github.com/solanabr/auditor-skill",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add auditor-skill", "env": [] },
"license": "MIT",
"maintainer": "solanabr",
"signal": { "stars": 58, "last_commit": "2026-09-29", "reputability": "org" },
"default_installed": true,
"safety": "seven executables in its own tree, plus a git hook, and a vendored submodule that does not ship. The executables: install.sh; scripts/check-attribution.sh, check-corpus.sh and check-hygiene.sh (the repo's own CI gates); scripts/report-to-pdf.sh (pandoc); templates/poc/run.sh (builds and runs a proof-of-concept against the audited program); and scripts/hooks/commit-msg, a git hook that rejects AI-attribution trailers — it is code that would run on every commit, but it sits in scripts/hooks/ and nothing installs it, so it fires only if someone points core.hooksPath at it. install.sh is for standalone installs and no kit installer runs it: it git clones the repo from github.com into .claude/skills/auditor-skill (or any path you pass, including ~/.claude/skills for every project), documents a curl | bash route, and then runs submodule update --init --recursive — which is exactly how the vendor tree below would arrive. That vendor/trailofbits submodule is recorded in vendored and is fetched by no kit installer, so no copy of it reaches a project; the pack tests for it and falls back when absent. No hooks.json, no .claude/settings.json, no .mcp.json, no symlinks, no package.json and so no install lifecycle hook, no nested .claude/skills/, no bidi or invisible Unicode. It ships AGENTS.md at its root, the nested instruction file an agent working inside the pack directory picks up. Its eval, child_process, spawn and 169.254.169.254 hits are all audit material — command-injection and SSRF entries in known-vectors/ and the language-safety checklists — not code. Its tool-execution steps run static analysis, fuzzers and coverage on the audited repo, so run them on code you intend to build.",
"tags": ["security", "audit", "solana", "anchor", "pinocchio", "supply-chain", "owasp"]
},
{
"id": "colosseum",
"name": "Colosseum Copilot",
"type": "skill",
"tier": "core",
"domain": "go-to-market",
"path": ".claude/skills/ext/colosseum",
"commit": "17e9bf07759d25fcec9cc3a751a081e448bc602c",
"description": "Colosseum Copilot 2.0: startup research, idea validation and competitive landscape from Colosseum's hackathon archives, plus feedback on the user's own submissions. Read by the solana-researcher and solana-architect agents and by the idea-sprint and hackathon skills. Sign in once per machine with the helper (Node 20+): npx @colosseum-org/copilot-connect login (add --device where no browser can open); npx @colosseum-org/copilot-connect status checks the saved connection, logout removes local credentials, revoke ends the connection. The helper keeps the credential in the OS store; no environment variable holds it.",
"triggers": ["Colosseum", "hackathon archives", "idea validation", "competitive research"],
"source": "https://github.com/ColosseumOrg/colosseum-copilot",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add colosseum", "env": [] },
"license": "proprietary",
"maintainer": "ColosseumOrg",
"signal": { "stars": 13, "last_commit": "2026-10-01", "reputability": "official" },
"default_installed": true,
"safety": "caution: queries Colosseum's API as the signed-in user and runs npx @colosseum-org/copilot-connect, an unpinned npm package, to hold the credential. COLOSSEUM_COPILOT_PAT and any api/v1 base are leftovers from the superseded v1 auth; v1 tokens stop working 2026-10-28. The skill shares each answer with Colosseum when the account has session sharing on. Mechanically the smallest core pack and inert: 16 files, no executables, no hooks.json, no .claude/settings.json, no .mcp.json, no symlinks, no package.json and so no install lifecycle hook, no .gitmodules of its own, no nested .claude/skills/, no root CLAUDE.md or AGENTS.md, no bidi or invisible Unicode. The egress is the copilot API the skill exists to call, and FRAMES_API_KEY in its text is a third-party key it names rather than reads.",
"tags": ["hackathon", "research", "ideas", "official", "proprietary"]
},
{
"id": "anthropic-skills",
"name": "Anthropic Skills (Apache-2.0 subset)",
"type": "skill",
"tier": "core",
"domain": "frontend-design",
"path": ".claude/skills",
"commit": "8a1541c4a3ffa5a20a5a91de0dcf3f0bab1d1ef4",
"skills": ["frontend-design", "webapp-testing"],
"description": "Anthropic's frontend-design (distinctive UI direction) and webapp-testing (Playwright tests of a local web app), installed as top-level skills: .claude/skills/<name>/ in a default install (Claude Code, Grok Build), or .agents/skills/<name>/ with install.sh --agents (Codex, Cursor, Copilot and other Agent Skills clients). One install serves one of the two families. The only core pack that is not a submodule: skills.sh fetches these two folders from anthropics/skills at the pinned commit and copies them unchanged, LICENSE.txt included, so a default install reaches a second host and a fetch that fails leaves the pack absent with a retry line rather than stopping the install.",
"triggers": ["UI design direction", "Playwright tests of a local web app"],
"source": "https://github.com/anthropics/skills",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add anthropic-skills", "env": [] },
"license": "Apache-2.0",
"maintainer": "anthropics",
"signal": { "stars": 179151, "last_commit": "2026-09-28", "reputability": "official" },
"default_installed": true,
"safety": "caution: webapp-testing's scripts/with_server.py runs the server commands it is given through a shell — the pack's one executable; the skills' guidance is agent-neutral, and the two folders carry nothing else from the mechanical set (no hooks.json, no .claude/settings.json, no .mcp.json, no symlinks, no package.json, no bidi or invisible Unicode, no credential read). The one core pack with a standing session cost: it installs top-level, where Claude Code auto-discovers it, so its two descriptions (408 characters) are listed every session and every subagent — about 102 tokens. The ext/ core packs cost nothing by comparison, but not because ext/ is exempt: Claude Code discovers any directory named .claude/skills at any depth, so a pack costs nothing standing only while it keeps its own skills somewhere else, which is true of every pinned pack except expo (see its entry). install.sh and skills.sh delete a pack's own .claude/ from the copy that lands in a project (.claude/bin/_pack_strip.sh), and tests/test_pack_load_surfaces.sh records which packs ship one, so a pack that gains one fails CI rather than shipping.",
"tags": ["frontend", "design", "playwright", "testing", "cross-agent", "official"]
},
{
"id": "qedgen",
"name": "QEDGen Solana Skills",
"type": "skill",
"tier": "extension",
"domain": "security",
"path": ".claude/skills/ext/qedgen",
"commit": "2c2c837175f6aaf0623820009b4a95371c204a97",
"description": "Formal verification for Solana programs: Lean 4 proofs of program invariants.",
"triggers": ["formal verification", "Lean 4", "invariant proofs"],
"source": "https://github.com/QEDGen/solana-skills",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add qedgen", "env": ["MISTRAL_API_KEY", "ARISTOTLE_API_KEY"] },
"license": "MIT",
"maintainer": "QEDGen",
"signal": { "stars": 68, "last_commit": "2026-09-25", "reputability": "org" },
"default_installed": false,
"safety": "twenty-one executables, one of which fetches a binary over the network, and the largest pack on disk (about 11 MB). install.sh is for standalone installs and no kit installer runs it: it curl's a release asset from github.com/QEDGen/solana-skills with --proto '=https', verifies it against a .sha256 fetched the same way, refuses to install an unverified or non-running candidate, falls back to a cargo build from source, and only symlinks the CLI into a directory you name with --link-dir, refusing to replace anything already there. skills/qedgen/install.sh is a second copy of it, tools/qedgen and skills/qedgen/tools/qedgen are shell wrappers around the CLI, and the remaining sixteen are the repo's own check-*, release-gate and preflight scripts plus one Claude Code hook script (integrations/qedgen-auditor-hooks/auditor-thinking-budget.sh) that no config here wires. Needs the qedgen CLI, plus MISTRAL_API_KEY for fill-sorry and generate or ARISTOTLE_API_KEY for the aristotle commands. No hooks.json, no .claude/settings.json, no .mcp.json, no symlinks, no package.json and so no install lifecycle hook, no .gitmodules of its own, no nested .claude/skills/. One U+200B sits in front of six code fences in skills/qedgen-auditor/references/report-and-grading.md — a Markdown artifact, not payload. It ships CLAUDE.md at its root, the nested instruction file an agent working inside the pack directory picks up, the same way it would an AGENTS.md there.",
"tags": ["formal-verification", "lean4", "security"]
},
{
"id": "solana-fuzz",
"name": "Solana Fuzz Skill",
"type": "skill",
"tier": "extension",
"domain": "testing-qa",
"path": ".claude/skills/ext/solana-fuzz",
"commit": "ff594b8050b2daaf3bce207f161492c2c44a667a",
"description": "The Trident harness API at a pinned version, plus the flag that makes a fuzz run mean anything. Two things the kit does not already have: the v0.12.0 call surface written out (#[flow_executor], #[init]/#[flow]/#[end], process_transaction, get_account_with_type and its discriminator offset, PDA seeds in fuzz_accounts) so an agent stops authoring Trident calls from memory against an API that moves every minor version; and the --with-exit-code trap, where a failing invariant is swallowed in parallel mode and the run still exits 0, so a green CI job proves nothing. No other pinned pack mentions that flag, and /test-rust and /audit-solana both print trident fuzz run without it. It layers under auditor-skill rather than overlapping: that pack already calls Trident the default primary harness and ships 19 invariants phrased as Trident post-conditions, deciding what to assert; this one is how to build the harness that asserts it. Ships one runnable Anchor example that goes red on a planted wrapping_sub bug and green on the fix, pins Trident 0.12.0 stable over the 0.13 RC, and marks its own regression and trident compare surface as unverified.",
"triggers": ["Trident", "fuzzing", "fuzz test", "invariant testing", "property testing"],
"source": "https://github.com/belumume/solana-fuzz",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add solana-fuzz", "env": [] },
"license": "MIT",
"maintainer": "belumume",
"signal": { "stars": 3, "last_commit": "2026-08-12", "reputability": "individual" },
"default_installed": false,
"safety": "Two executables and no env keys, no nested submodules; it only writes test code. One is install.sh, for standalone installs, which no kit installer runs: it copies its own skill folder into ${SKILLS_DIR:-~/.claude/skills} with cp -R and makes no network call. The other is .githooks/commit-msg, a POSIX sh commit-message hook the pack's own root CLAUDE.md asks an agent to enable with git config core.hooksPath .githooks; it strips Claude attribution trailers and claude.ai session links from the message with grep/sed/awk and makes no network call. The kit firewall denies that hooksPath write at every tier, so the hook stays inert in a Claude Code session; only an --agents install without the guard would run it. Nothing else from the mechanical set — no hooks.json, no .claude/settings.json, no .mcp.json, no symlinks, no package.json and so no install lifecycle hook, no nested .claude/skills/, no AGENTS.md, no bidi or invisible Unicode. It does ship a root CLAUDE.md: a no-Claude-attribution policy for its own repo that asks an agent to enable the commit-msg hook above and to merge without squash; benign and scoped to the pack, and the hooksPath line it names the kit firewall denies anyway. The Trident CLI it drives is a separate cargo install trident-cli --locked, and a fuzz run builds and executes the program under test, so run it on code you intend to build. Pinned pre-fix: the worked example still pins anchor-lang 0.31 and uses the pre-1.0 space = 8 + INIT_SPACE form (belumume/solana-fuzz#1, fix open in #2). That is the example program only -- the harness API this pack is pinned for is unaffected, since trident-tests/ depends on borsh and trident-fuzz and not on anchor-lang.",
"tags": ["testing", "fuzzing", "trident", "invariants", "solana", "anchor"]
},
{
"id": "sendai",
"name": "SendAI Skills",
"type": "skill",
"tier": "extension",
"domain": "solana-defi",
"path": ".claude/skills/ext/sendai",
"commit": "06b36b8b420e705ffd57b29e463f01093959d8d4",
"description": "SendAI's protocol skills: Raydium, Orca, Meteora, Kamino, marginfi, Sanctum, Pyth, Switchboard, Squads, pump.fun, deBridge, LI.FI, Light Protocol and more, plus the web3.js to Kit migration skills. Its jupiter, metaplex, helius and magicblock folders are older than the official packs.",
"triggers": ["DeFi protocols", "Raydium", "Orca", "Meteora", "Kamino", "marginfi", "Sanctum", "Pyth", "Switchboard", "Squads", "pump.fun", "bridging", "web3.js to Kit migration"],
"source": "https://github.com/sendaifun/skills",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add sendai", "env": [] },
"license": "Apache-2.0",
"maintainer": "sendaifun",
"signal": { "stars": 128, "last_commit": "2026-07-31", "reputability": "org" },
"default_installed": false,
"safety": "four executables and one MCP entry; community skills, so prefer a protocol's official pack where one exists. The executables are per-skill helpers, none of them an installer: animation-reverse-engineering/scripts/dissect.sh, pinocchio-development/scripts/scaffold-program.sh, and solana-kit-migration/scripts/{analyze-migration.sh,detect-patterns.js}, which read a codebase and scaffold or report. skills/arcium/mcp.json names one remote entry, https://docs.arcium.com/mcp, and a pinned pack's MCP config does not take effect until you add it yourself. No hooks.json, no .claude/settings.json, no symlinks, no package.json and so no install lifecycle hook, no .gitmodules of its own, no nested .claude/skills/, no root CLAUDE.md or AGENTS.md, no bidi or invisible Unicode. The eval, new Function and exec hits are vulnhunter's detection patterns — the strings it tells an auditor to grep for — except skills/ct-alpha/install.ts, which spawns a package manager for its own setup. Five surfpool and arcium reference files tell you to pipe an installer into a shell; those are commands for you to run. Its env-var list is wide because it documents other vendors' keys (BIRDEYE_API_KEY, HELIUS_API_KEY, SOLANA_PRIVATE_KEY and the rest) in integration examples rather than reading any of them.",
"tags": ["solana", "defi", "protocols", "oracles", "bridges"]
},
{
"id": "jupiter",
"name": "Jupiter Agent Skills",
"type": "skill",
"tier": "extension",
"domain": "solana-defi",
"path": ".claude/skills/ext/jupiter",
"commit": "a2211e3f6a7caa03310c7a9a8d816e723b0bdc5f",
"description": "Jupiter's official skills: integrating-jupiter (swap, lend, perps, trigger, recurring), jupiter-lend, jupiter-swap-migration, jupiter-vrfd.",
"triggers": ["Jupiter", "token swaps", "Jupiter Lend", "perps", "trigger orders", "recurring orders"],
"source": "https://github.com/jup-ag/agent-skills",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add jupiter", "env": [] },
"license": "MIT",
"maintainer": "jup-ag",
"signal": { "stars": 84, "last_commit": "2026-07-07", "reputability": "official" },
"default_installed": false,
"safety": "two executables and two MCP entries, all plugin plumbing rather than skill content. scripts/install_plugin.sh registers the bundled integrate-jupiter plugin with a local Codex marketplace and scripts/sync_plugin_skills.sh copies the skill folders into it; neither is run by a kit installer. .plugins/integrate-jupiter/claude/.mcp.json and its codex sibling each name one remote entry, https://developers.jup.ag/docs/mcp, and a pinned pack's MCP config does not take effect until you add it yourself. No hooks.json, no .claude/settings.json, no symlinks, no package.json and so no install lifecycle hook, no .gitmodules of its own, no nested .claude/skills/, no root CLAUDE.md or AGENTS.md, no bidi or invisible Unicode, no eval, atob, new Function, createRequire or child_process. WALLET_PRIVATE_KEY and JUPITER_API_KEY appear as documented env names in examples, not as reads.",
"tags": ["solana", "defi", "jupiter", "swap", "official"]
},
{
"id": "position-manager-skill",
"name": "Position Manager Skill",
"type": "skill",
"tier": "extension",
"domain": "solana-defi",
"path": ".claude/skills/ext/position-manager-skill",
"commit": "4885751c627a1db038e4aa0a513d248895fe6a0c",
"description": "Concentrated-liquidity LP lifecycle across Orca Whirlpools, Raydium CLMM, Meteora DLMM and DAMM v2, and Kamino vaults: position and unclaimed-fee tracking, realtime out-of-range alerts over gRPC or websocket rather than polling, IL-vs-HODL and fee-APR backtesting with break-even range width, cross-DEX yield comparison, simulate-first rebalancing behind a typed human confirmation, and a FIFO/HIFO/SPECID tax-lot ledger. One SKILL.md routes to ten reference leaves (open exactly one per task) over TypeScript templates and per-venue adapters; it delegates Anchor, IDL and transaction plumbing to solana-dev and swaps to the jupiter skill. Scaffolding and knowledge, not a hosted or auto-running bot.",
"triggers": ["concentrated liquidity", "LP positions", "out-of-range alerts", "impermanent loss", "rebalancing a position", "LP tax lots"],
"source": "https://github.com/solanabr/position-manager-skill",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add position-manager-skill", "env": [] },
"license": "MIT",
"maintainer": "solanabr",
"signal": { "stars": 2, "last_commit": "2026-07-10", "reputability": "org" },
"default_installed": false,
"safety": "caution: its templates build and send transactions that move funds. The skill's own rails default to DRY_RUN=true and REQUIRE_CONFIRM=true with slippage, max-notional, max-position and daily-loss caps plus a kill switch, and it requires a typed confirmation before any signing path; keep those defaults. Needs an RPC endpoint, and a funded keypair for writes. Mechanically it ships no executables — the install.sh and install-custom.sh at its repo root are shipped without the exec bit and are for standalone installs; the kit vendors the pack instead, so there is nothing to run, but if you run one by hand it rm -rf's and replaces ~/.claude/skills/position-manager and ~/.codex/skills, fetches solana-dev-skill from github.com, and copies its agents/, commands/ and rules/ into ~/.claude/ for every project. Nothing else from the mechanical set: no hooks.json, no .claude/settings.json, no .mcp.json, no symlinks, no package.json and so no install lifecycle hook, no .gitmodules of its own, no nested .claude/skills/, no root CLAUDE.md or AGENTS.md, no bidi or invisible Unicode, and no eval, atob, new Function, createRequire or child_process. Smallest of the kit's packs on disk (about 356 KB) and ~88 tokens of session listing.",
"tags": ["solana", "defi", "clmm", "liquidity", "orca", "raydium", "meteora", "kamino"]
},
{
"id": "metaplex",
"name": "Metaplex Skill",
"type": "skill",
"tier": "extension",
"domain": "solana-nft",
"path": ".claude/skills/ext/metaplex",
"commit": "23a17cbe848c0e38aa29ae565fabdcf8fdce7d4b",
"description": "Metaplex's official skill: Core, Token Metadata, Bubblegum compressed NFTs, Candy Machine, Genesis launches, Umi.",
"triggers": ["NFTs", "Metaplex Core", "Token Metadata", "Bubblegum", "Candy Machine", "Umi"],
"source": "https://github.com/metaplex-foundation/skill",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add metaplex", "env": [] },
"license": "Apache-2.0",
"maintainer": "metaplex-foundation",
"signal": { "stars": 21, "last_commit": "2026-07-21", "reputability": "official" },
"default_installed": false,
"safety": "inert on every mechanical category: 26 files, zero executables, no hooks.json, no .claude/settings.json, no .mcp.json, no symlinks, no package.json and so no install lifecycle hook, no .gitmodules of its own, no nested .claude/skills/, no root CLAUDE.md or AGENTS.md, no bidi or invisible Unicode, and no eval, atob, new Function, createRequire or child_process. Markdown only. Hosts named in its text are metaplex.com, arweave.net and gateway.irys.xyz — upload and read targets its own guidance uses — and its keypair and id.json mentions are prose about the reader's own wallet rather than reads.",
"tags": ["solana", "nft", "metaplex", "cnft", "official"]
},
{
"id": "magicblock",
"name": "MagicBlock Dev Skill",
"type": "skill",
"tier": "extension",
"domain": "solana-infra",
"path": ".claude/skills/ext/magicblock",
"commit": "de3c62bcf39b72a8b1ffd7e65c434be3c9f0edc8",
"description": "MagicBlock's official skill: Ephemeral Rollups with delegated state, ER and PER architecture and settlement, private payments, session keys, VRF, cranks. Replaces the sendai magicblock folder, which pins ephemeral-rollups-sdk 0.6.5 (this pack covers 0.16.2).",
"triggers": ["MagicBlock", "Ephemeral Rollups", "delegated state", "real-time games", "on-chain randomness"],
"source": "https://github.com/magicblock-labs/magicblock-dev-skill",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add magicblock", "env": [] },
"license": "MIT",
"maintainer": "magicblock-labs",
"signal": { "stars": 9, "last_commit": "2026-09-24", "reputability": "official" },
"default_installed": false,
"safety": "two executables, one of them a global installer. install.sh is for standalone installs and no kit installer runs it: it writes the skill into ${CLAUDE_HOME:-$HOME/.claude}, ${CODEX_HOME:-$HOME/.codex}, or the Cursor, Windsurf, Cline and Continue equivalents depending on the flags, and can drop an AGENTS.md at a project root; build.sh assembles the dist/ bundle it installs. No hooks.json, no .claude/settings.json, no .mcp.json, no symlinks, no package.json and so no install lifecycle hook, no .gitmodules of its own, no nested .claude/skills/, no root CLAUDE.md or AGENTS.md, no bidi or invisible Unicode, and no eval, atob, new Function, createRequire or child_process. Egress in the text is MagicBlock's own routers and RPC hosts.",
"tags": ["solana", "ephemeral-rollups", "gaming", "real-time", "official"]
},
{
"id": "helius",
"name": "Helius Core AI",
"type": "skill",
"tier": "extension",
"domain": "solana-infra",
"path": ".claude/skills/ext/helius",
"commit": "5f1da9384b3f61560973fd88fe8003b0eeabce4b",
"description": "Helius's official skills: RPC, DAS, webhooks, WebSockets, Laserstream, Sender, priority fees, SVM internals and partner integrations. The kit's Helius MCP server answers doc questions without it.",
"triggers": ["Helius", "DAS", "webhooks", "Laserstream", "Sender", "priority fees", "SVM internals"],
"source": "https://github.com/helius-labs/core-ai",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add helius", "env": [] },
"license": "MIT",
"maintainer": "helius-labs",
"signal": { "stars": 28, "last_commit": "2026-08-31", "reputability": "official" },
"default_installed": false,
"safety": "two executables, two MCP entries and two root instruction files. helius-skills/helius-phantom/install.sh is for standalone installs and no kit installer runs it: it defaults to $HOME/.claude/skills/helius-phantom, with --project and --path for the other two targets. helius-cli/scripts/verify-envelope.sh is a release-signing check. helius-plugin/.mcp.json and helius-cursor/.mcp.json are identical and stdio rather than remote: npx helius-mcp@2.2.0, the vendor's own package at a pinned version, and a pinned pack's MCP config does not take effect until you add it yourself. Its two package.json files declare prepublishOnly only, which runs on npm publish and never on install. No hooks.json, no .claude/settings.json, no symlinks, no .gitmodules of its own, no nested .claude/skills/, no bidi or invisible Unicode. The execSync and spawn hits are helius-cli's own self-updater and browser launcher, and the atob hits are base64 decoding in frontend sample code. It ships both AGENTS.md and CLAUDE.md at its root as real files — the nested instruction files an agent working inside the pack directory picks up. HELIUS_API_KEY is the env var throughout; the other key names are documented third-party vendors' (Jupiter, OKX, DFlow), not reads.",
"tags": ["solana", "rpc", "das", "webhooks", "official"]
},
{
"id": "alchemy",
"name": "Alchemy Skills",
"type": "skill",
"tier": "extension",
"domain": "solana-infra",
"path": ".claude/skills/ext/alchemy",
"commit": "4caf3e555c3af8dbb9e5b540aed558aef986d885",
"description": "Alchemy's official skills: the Alchemy API, CLI and MCP server, with Solana RPC, DAS, Yellowstone gRPC and WebSocket references, plus a keyless x402 gateway for app code.",
"triggers": ["Alchemy APIs", "Yellowstone gRPC", "x402 gateway"],
"source": "https://github.com/alchemyplatform/skills",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add alchemy", "env": ["ALCHEMY_API_KEY"] },
"license": "MIT",
"maintainer": "alchemyplatform",
"signal": { "stars": 49, "last_commit": "2026-08-26", "reputability": "official" },
"default_installed": false,
"safety": "two executables, both plugin plumbing, and each skill ships twice (skills/ and plugins/alchemy/skills/). scripts/install_codex_plugin.py copies the bundled plugin into a user's home-local Codex setup and plugins/alchemy/scripts/sync_from_catalog.py regenerates the duplicate tree; no kit installer runs either. No hooks.json, no .claude/settings.json, no .mcp.json, no symlinks, no package.json and so no install lifecycle hook, no .gitmodules of its own, no nested .claude/skills/, no root CLAUDE.md or AGENTS.md, no bidi or invisible Unicode, and no eval, atob, new Function, createRequire or child_process. ALCHEMY_API_KEY and the webhook secrets are documented env names in examples rather than reads, and the hosts are Alchemy's own.",
"tags": ["rpc", "das", "grpc", "multi-chain", "official"]
},
{
"id": "solana-game",
"name": "Solana Game Skill",
"type": "skill",
"tier": "extension",
"domain": "gaming",
"path": ".claude/skills/ext/solana-game",
"commit": "762b3b28e62b4c6cabc01ae179d316958a852a07",
"vendored": { "skill/solana-dev-skill": "dcc633a384c033bb1ee20c26325caec6c611ecae" },
"description": "Superteam Brazil's game skill: Solana.Unity-SDK, C# patterns, game architecture, PlaySolana (PSG1, PlayDex, PlayID), mobile game patterns. Read by the game-architect and unity-engineer agents.",
"triggers": ["Unity", "C#", "games", "PlaySolana", "PSG1"],
"source": "https://github.com/solanabr/solana-game-skill",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add solana-game", "env": [] },
"license": "MIT",
"maintainer": "solanabr",
"signal": { "stars": 9, "last_commit": "2026-01-30", "reputability": "org" },
"default_installed": false,
"safety": "two executables, and the one named install.sh overwrites a global instruction file. It is for standalone installs and no kit installer runs it, but if you run it yourself it copies the pack's own CLAUDE.md over $HOME/.claude/CLAUDE.md — your global agent instructions, in every project — after moving the old one aside to CLAUDE.md.backup, and rm -rf's and replaces $HOME/.claude/skills/solana-dev and $HOME/.claude/skills/solana-game, fetching the first from github.com. install-custom.sh is the same flow with a chosen target. It carries an older copy of solana-dev-skill as a nested submodule, recorded in vendored and fetched by no kit installer. No hooks.json, no .claude/settings.json, no .mcp.json, no symlinks, no package.json and so no install lifecycle hook, no nested .claude/skills/, no bidi or invisible Unicode, and no eval, atob, new Function, createRequire or child_process. It ships CLAUDE.md at its root — the nested instruction file an agent picks up inside the pack directory, the same way it would an AGENTS.md there, and also the file that installer would publish globally.",
"tags": ["gaming", "unity", "csharp", "playsolana"]
},
{
"id": "solana-mobile",
"name": "Solana Mobile Skills",
"type": "skill",
"tier": "extension",
"domain": "mobile",
"path": ".claude/skills/ext/solana-mobile",
"commit": "8664b5be49d16b6962129b53def20f6dfa99443d",
"description": "Solana Mobile's official skills: React Native and Expo scaffolding, Mobile Wallet Adapter 2.0, Seeker Genesis Token and .skr domains, dApp Store publishing. Read by the mobile-engineer agent.",
"triggers": ["React Native", "Expo", "Mobile Wallet Adapter", "Seeker", "dApp Store"],
"source": "https://github.com/solana-mobile/solana-mobile-skills",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add solana-mobile", "env": [] },
"license": "Apache-2.0",
"maintainer": "solana-mobile",
"signal": { "stars": 8, "last_commit": "2026-09-24", "reputability": "official" },
"default_installed": false,
"safety": "three executables, all repo CI: scripts/check-links.sh, scripts/validate-skill.py and scripts/validate-skills.sh, which lint the pack's own frontmatter and links. No installer of any kind. No hooks.json, no .claude/settings.json, no .mcp.json, no symlinks, no package.json and so no install lifecycle hook, no .gitmodules of its own, no nested .claude/skills/, no bidi or invisible Unicode. The one createRequire hit is a Node pattern in the seeker-domains sample server. It ships AGENTS.md at its root, the nested instruction file an agent working inside the pack directory picks up. DAPP_STORE_API_KEY and PRIVY_APP_SECRET are documented env names in publishing examples rather than reads.",
"tags": ["mobile", "react-native", "mwa", "seeker", "official"]
},
{
"id": "cloudflare",
"name": "Cloudflare Skills",
"type": "skill",
"tier": "extension",
"domain": "infra",
"path": ".claude/skills/ext/cloudflare",
"commit": "cd687a9adc046b8fcbea6b747f9008f0e112ccee",
"description": "Cloudflare's official skills: Workers best practices, wrangler, Durable Objects, Agents SDK, sandboxes. Read by the devops-engineer agent for Workers code.",
"triggers": ["Cloudflare Workers", "wrangler", "Durable Objects", "Agents SDK"],
"source": "https://github.com/cloudflare/skills",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add cloudflare", "env": [] },
"license": "Apache-2.0",
"maintainer": "cloudflare",
"signal": { "stars": 2907, "last_commit": "2026-09-22", "reputability": "official" },
"default_installed": false,
"safety": "four executables and two MCP entries. The executables all sit under skills/turnstile-spin/scripts — auth-probe.sh, widget-create.sh, validate.sh and persist-skill.sh — and drive wrangler and the Cloudflare API against whichever account your ambient login holds; widget-create.sh creates a real Turnstile widget, and persist-skill.sh writes a skill bundle into the current project. Each one unsets CLOUDFLARE_API_TOKEN, CF_API_TOKEN, GITHUB_TOKEN, NPM_TOKEN and the rest before doing anything, which is deliberate hygiene rather than a credential read. .mcp.json and mcp.json are the same single remote entry, https://mcp.cloudflare.com/mcp, and a pinned pack's MCP config does not take effect until you add it yourself. No hooks.json, no .claude/settings.json, no symlinks, no package.json and so no install lifecycle hook, no .gitmodules of its own, no nested .claude/skills/, no root CLAUDE.md or AGENTS.md, no bidi or invisible Unicode. The eval, atob and child_process hits are workerd API documentation.",
"tags": ["infra", "workers", "edge", "official"]
},
{
"id": "vercel",
"name": "Vercel Agent Skills",
"type": "skill",
"tier": "extension",
"domain": "frontend-design",
"path": ".claude/skills/ext/vercel",
"commit": "1a2c79d4e6c4a2283dc2c614b6538cbe1ed5ca62",
"description": "Vercel's agent skills: React and Next.js performance, composition patterns, web design guidelines, deploying to Vercel.",
"triggers": ["Vercel", "Next.js", "React performance", "web design review"],
"source": "https://github.com/vercel-labs/agent-skills",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add vercel", "env": [] },
"license": "MIT",
"maintainer": "vercel-labs",
"signal": { "stars": 31526, "last_commit": "2026-08-28", "reputability": "official" },
"default_installed": false,
"safety": "nine executables, one of which uploads your project to a remote. skills/deploy-to-vercel/resources/deploy.sh tars the directory you point it at and POSTs it to https://claude-skills-deploy.vercel.com/api/deploy, an anonymous claimable-deploy endpoint on the vendor's own domain, returning a preview URL and a claim URL; it excludes .env, .env.*, .git and node_modules from the tarball, so a plain run does not ship your secrets, but it does put your source on someone else's host. deploy-codex.sh is the same flow for Codex, and the other seven are skills/vercel-optimize/scripts/*.mjs, which scan a codebase and build docs locally. It ships one symlink, CLAUDE.md -> AGENTS.md at the pack root, resolving to the real AGENTS.md beside it; see the redis entry for exactly what each copy path does with it. No hooks.json, no .claude/settings.json, no .mcp.json, no package.json install lifecycle hook, no .gitmodules of its own, no nested .claude/skills/, no bidi or invisible Unicode. The child_process and exec hits are its own test suite under packages/vercel-optimize-tests.",
"tags": ["frontend", "nextjs", "react", "deploy", "official"]
},
{
"id": "get-shit-pretty",
"name": "Get Shit Pretty (GSP)",
"type": "skill",
"tier": "extension",
"domain": "frontend-design",
"path": ".claude/skills/ext/get-shit-pretty",
"commit": "5754ce8c44cd9cc9d9207e25955a1a89cf875bb8",
"description": "Design engineering for a dApp's surface: a brand pipeline (strategy, identity, guidelines) and a project pipeline (brief, research, design, critique, build, review) across 45 skills and 13 scoped sub-agents, emitting shadcn registry themes, Tailwind tokens, a STYLE.md and WCAG 2.2 audits. 35 style presets. Not Solana-specific — it designs the client, not the program.",
"triggers": ["brand identity or a design system for the app", "shadcn theme or design tokens", "make the UI less generic", "accessibility audit of a screen", "design critique before building"],
"source": "https://github.com/jubscodes/get-shit-pretty",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add get-shit-pretty", "env": [] },
"license": "MIT",
"maintainer": "jubscodes",
"signal": { "stars": 55, "last_commit": "2026-06-06", "reputability": "individual" },
"default_installed": false,
"safety": "fifteen executables, and clean at this pin on every other checklist item, verified across all 1,657 blobs in its history and not just the checkout: no eval, atob, new Function or createRequire; no install lifecycle hook in package.json, ever; zero occurrences of keypair, wallet, private-key, mnemonic or credential terms anywhere in the repo; no invisible or bidi Unicode; no whitespace-padded line; no base64 payload; no symlinks; no .claude/settings.json; no .gitmodules of its own; no nested .claude/skills/; every long line is prose or a Tailwind class string. Zero production dependencies. The fifteen are install.sh and bin/install.js, two statusline and preview helpers under gsp/skills/*/bin/, and eleven dev and CI scripts under scripts/ and dev/scripts/; child_process appears in two statusline scripts only, argv-form with fixed arguments, and bin/install.js uses none. Its only egress is `npm view get-shit-pretty version` once per 24h, and the only IP literal is 127.0.0.1 in a loopback preset server that self-exits after 60s. Pinning it is inert for the config files: the kit copies ext/ packs as plain files, fetches no pack submodules and never reads a pack's .mcp.json or hooks.json, so neither takes effect in a session. Three things to know before wiring any of it by hand. (1) Its root .mcp.json declares a local stdio MCP server, `npx -y @anthropic/mcp-figma` carrying FIGMA_ACCESS_TOKEN, and that package does not exist: the whole @anthropic npm scope is unclaimed (Anthropic's is @anthropic-ai), so the name is squattable and a local MCP server runs outside the OS sandbox. Do not copy that file to a project root. (2) gsp/hooks/hooks.json wires PreToolUse on every Edit/Write, SessionStart on compact and twelve SubagentStop matchers to bash scripts/*.sh; those scripts are local, read-only over .design/ and parse stdin with jq, but they run unattended once wired. (3) dev/scripts/token-proxy.py is a mitmproxy addon that intercepts api.anthropic.com to count tokens; it records only usage counters to a local jsonl, transmits nothing, and needs both mitmproxy and an HTTPS_PROXY you set yourself. Correcting the earlier add-on note: bin/install.js does not merge hooks into settings.json — it writes settings.statusLine only and never references hooks.json or .mcp.json, so the pack's own installer does not wire its hooks either, despite its README advertising them. Its install.sh is stale (it copies commands/ and agents/ paths that no longer exist, under set -e) and targets ~/.claude globally, writing into ~/.claude/commands/gsp and ~/.claude/agents; nothing in the kit runs it. It ships CLAUDE.md at its root, the nested instruction file an agent working inside the pack directory picks up, the same way it would an AGENTS.md there. A pin vendors the author's whole monorepo, ~4.6 MB over 571 files, of which the gsp/ payload is 1.6 MB and the rest is their marketing site, their own .design artifacts, dev tooling and nine commercial Pangram Pangram woff2 fonts. In --agents installs Codex and opencode load all 45 nested SKILL.md files.",
"tags": ["design", "branding", "design-systems", "ui-polish", "frontend", "accessibility", "shadcn", "hooks"]
},
{
"id": "obsidian-skills",
"name": "Obsidian Skills",
"type": "skill",
"tier": "extension",
"domain": "knowledge-management",
"path": ".claude/skills/ext/obsidian-skills",
"commit": "3ccff5338ea700537839b21900aa5358a0402c98",
"description": "Reading and writing an Obsidian vault as the project's notes layer: Obsidian-flavored Markdown (callouts, embeds, properties), Bases (.base query views plus their function reference), JSON Canvas (.canvas), Knap, the `obsidian` CLI for vault operations and plugin development, and Defuddle for turning a web page into clean Markdown. Not Solana-specific — it manages notes and research, not programs.",
"triggers": ["an Obsidian vault or .base/.canvas files in one", "research notes or a knowledge base for the project", "Obsidian-flavored Markdown with callouts or embeds or properties", "a Bases query view or a JSON Canvas board", "extracting a web page to clean Markdown"],
"source": "https://github.com/kepano/obsidian-skills",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add obsidian-skills", "env": [] },
"license": "MIT",
"maintainer": "kepano",
"signal": { "stars": 49153, "last_commit": "2026-09-15", "reputability": "individual" },
"default_installed": false,
"safety": "clean at this pin on every checklist item, verified across the full history (49 commits, 50 blobs, largest blob 17.9 KB) and not just the checkout: zero symlinks, zero submodules, zero executables; no eval, atob, new Function or createRequire; no child_process, subprocess or os.system; no invisible or bidi Unicode; no line over 500 characters and no whitespace padding; no raw IP; no credential-shaped string. The pack ships no executable code at all — 15 tracked files, 12 Markdown plus two .claude-plugin manifests and a license — and every path that has ever existed in its history ended in .md or .json, so there is no package.json (hence no preinstall/postinstall/prepare hook, ever), no hooks.json, no .mcp.json and no .claude/settings.json. Its plugin manifests declare no hooks and no MCP servers. Hardcoded hosts are obsidian.md, github.com, jsoncanvas.org, agentskills.io, stephango.com and example.com placeholders. Two behaviours the skills instruct an agent to perform — both documented first-party features, not hidden payloads, and neither triggered by pinning: (1) obsidian-cli/SKILL.md drives the `obsidian` CLI against a running Obsidian instance and documents `obsidian eval code=\"...\"`, which executes arbitrary JavaScript in the app context, alongside dev:dom, dev:screenshot, dev:console and CDP/debugger controls — note that a grep for `eval(` does not catch this form, and it reaches the user's live vault; (2) defuddle/SKILL.md tells the agent to `npm install -g defuddle` (the package does exist on npm) and then fetch arbitrary URLs through it. Pinning is inert — the kit copies ext/ packs as plain files and never reads a pack's manifests — but both run if an agent follows the skill, so treat this pack as vault-write and local-exec capable once installed.",
"tags": ["obsidian", "notes", "knowledge-management", "markdown", "canvas", "research", "pkm"]
},
{
"id": "huggingface-skills",
"name": "Hugging Face Skills",
"type": "skill",
"tier": "extension",
"domain": "ai-ml",
"path": ".claude/skills/ext/huggingface-skills",
"commit": "25910b3d0ba31a2ba102f901cdc9e008c278a69c",
"description": "Hugging Face's own 25 skills for building AI/LLM features: the hf CLI and Hub operations (repos, datasets, models, Spaces, jobs), Gradio and Spaces apps, ZeroGPU, LLM and vision fine-tuning, TRL and sentence-transformers training, LoRA space building, community evals, transformers.js, local models, Trackio run tracking, and SageMaker deployment. Not Solana-specific — useful when a project puts a model or an inference endpoint behind its program or client.",
"triggers": ["building an AI or LLM feature for the app", "fine-tuning or training a model", "a Gradio app or a Hugging Face Space", "pulling a model or dataset off the Hub", "deploying an inference endpoint"],
"source": "https://github.com/huggingface/skills",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add huggingface-skills", "env": [] },
"license": "Apache-2.0",
"maintainer": "huggingface",
"signal": { "stars": 11139, "last_commit": "2026-10-01", "reputability": "official" },
"default_installed": false,
"safety": "nineteen executables, and clean at this pin on the blocking checklist items, verified across the full history (448 commits, 1126 blobs) and not just the checkout: zero symlinks, zero submodules; no eval, atob, new Function or createRequire (the only .eval( hits are PyTorch model.eval()); no invisible or bidi Unicode; no package.json anywhere, so no preinstall/postinstall/prepare hook, ever; no hooks.json, no .claude/settings.json and no nested .claude/skills/; no credential-shaped string outside an .env.example placeholder; and every path that has ever existed in its history is text (md, py, json, yml, mdx, txt, sh, toml, tsx, jsonl, html, png) with no binary, archive or compiled blob at any commit. Long lines exist (about 100 over 500 characters, longest 1588) and every one checked is a Gradio API signature or a Markdown table, not padding; the longest whitespace runs are table alignment. This pack is machinery, not prose, and five things follow. (1) Of the nineteen executables, fifteen are Python scripts inside skill directories and the other four are repo tooling (scripts/publish.sh plus three .py), with 94 subprocess call sites in all; every one is argv-form against aws, git, make, hf or uv, and none builds a shell string from input. (2) skills/hf-cli/SKILL.md and train-sentence-transformers tell the agent to run `curl -LsSf https://hf.co/cli/install.sh | bash -s`, a pipe-to-shell installer, and other skills run `uvx hf-mem`, `uvx trl-jobs` and `npx -y @huggingface/hub` — every one of those package names was verified to exist on PyPI or npm, so none is an unclaimed typosquat, but they fetch and execute code at run time. (3) The hf-cloud-sagemaker-* skills act on real AWS: scripts/create_role.py calls iam create-role and put-role-policy, and hf-cloud-aws-context-discovery/SKILL.md instructs the agent that ~/.aws/config and ~/.aws/credentials are plain INI files it may read. The IAM templates it installs are genuinely least-privilege (S3 scoped to one bucket, ECR pull, CloudWatch logs scoped to /aws/sagemaker/*, trust limited to sagemaker.amazonaws.com), but the pack creates cloud identity and reads credential files when followed. (4) Its root .mcp.json declares one remote HTTP MCP server, https://huggingface.co/mcp?login — no package to typosquat, and inert from a pin since the kit never reads a pack's .mcp.json, but do not copy that file to a project root without deciding you want an authenticated HF MCP server. HF_TOKEN is the env var throughout, as expected. Also ships a stray 3-line fast-agent-log.jsonl debug log at the root, checked and credential-free. (5) In an --agents install its 25 nested SKILL.md files land inside the tree Codex and opencode scan, so they do count against the shared description budget: skills.sh select keeps any pack named with --with, the copy loop vendors it to .agents/skills/ext/huggingface-skills/, install.sh gitignores that exact path, and the installer's closing banner tells the user Codex and opencode read the skills in .agents/skills/. Verified by running install.sh --agents --with huggingface-skills against a stand-in pack carrying nested SKILL.md files: all of them arrived under .agents/skills/ext/, and no .claude/ was created. The skills/ext -prune in install.sh and update.sh excludes ext/ only from the .claude/-to-.agents/ path rewrite (agents_paths), not from installation or discovery. Codex splits one shared ~22 KB budget across every skill description it registers and truncates each to its share with no warning, so installing this pack in --agents mode shrinks what every other skill description gets.",
"tags": ["ai", "llm", "huggingface", "machine-learning", "gradio", "spaces", "fine-tuning", "inference"]
},
{
"id": "supabase",
"name": "Supabase Agent Skills",
"type": "skill",
"tier": "extension",
"domain": "data",
"path": ".claude/skills/ext/supabase",
"commit": "c9be0e931b7930f7d02126d04774d904c381e7d7",
"description": "Supabase's official skills for the off-chain half of a Solana app. The supabase skill covers Database, Auth, Edge Functions, Realtime, Storage, Vectors, Cron and Queues, supabase-js and @supabase/ssr in the usual frameworks, the CLI, migrations and declarative schemas, and reading logs. The supabase-postgres-best-practices skill is a 34-file Postgres reference that holds for Postgres anywhere: query performance and EXPLAIN plans, connection pooling and exhaustion, partitioning, RLS policy performance, missing foreign-key indexes, and SKIP LOCKED queue patterns — the one to read when a Solana indexer or webhook consumer is writing high-volume rows.",
"triggers": ["Supabase", "Postgres schema or migration", "row-level security policy", "slow query or connection exhaustion", "indexer or webhook database writes"],
"source": "https://github.com/supabase/agent-skills",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add supabase", "env": [] },
"license": "MIT",
"maintainer": "supabase",
"signal": { "stars": 2693, "last_commit": "2026-10-02", "reputability": "official" },
"default_installed": false,
"safety": "no executables at this pin: both skill folders are markdown only (40 .md files, no scripts or binaries), package.json has no install lifecycle hook, and no eval, atob, createRequire or new Function appears anywhere in the pinned ancestry. A new Function does exist on the unmerged feature/workflow-evals branch, which the pin does not reach. The root test/sanity.test.ts calls execSync to run npx skills add into a temp dir — a vendor dev harness, never run by an install. No hooks.json, no .claude/settings.json, no .gitmodules of its own, no nested .claude/skills/, no bidi or invisible Unicode. It ships one symlink, CLAUDE.md -> AGENTS.md at the pack root, resolving to the real AGENTS.md beside it; see the redis entry for exactly what each copy path does with it. Its root .mcp.json names one remote entry, mcp.supabase.com, and a pinned pack's MCP config does not take effect until you add it yourself. Two behaviours to know rather than defects. Standing egress: skills/supabase/SKILL.md line 16 tells the agent to fetch https://supabase.com/changelog.md whenever the skill loads, and supabase.com/docs/**.md at several decision points, so every Supabase task makes a vendor-host call; no credentials travel with it. Outbound publishing: references/skill-feedback.md can open a public GitHub issue on supabase/agent-skills from a conversation-derived template. Its step 1 is to ask permission, so it is gated, but it is a path that puts conversation content outside the machine, and this kit has been conservative about those",
"tags": ["data", "postgres", "backend", "supabase", "official"]
},
{
"id": "solana-new",
"name": "solana.new",
"type": "skill",
"tier": "extension",
"domain": "go-to-market",
"path": ".claude/skills/ext/solana-new",
"commit": "38a494a21a3b0673e2fd581623a7f80861abd83b",
"description": "SendAI and Superteam go-to-market references: marketing video, brand and frontend design, tokenomics, DefiLlama research, launch playbooks, and ecosystem catalogs in cli/data/. Read its reference files directly; its SKILL.md files open with telemetry preambles whose bash blocks the kit does not run.",
"triggers": ["go-to-market", "marketing video", "brand design", "tokenomics", "DefiLlama research", "ecosystem catalogs"],
"source": "https://github.com/sendaifun/solana-new",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add solana-new", "env": [] },
"license": "MIT",
"maintainer": "sendaifun",
"signal": { "stars": 176, "last_commit": "2026-07-31", "reputability": "org" },
"default_installed": false,
"safety": "caution: seven executables, a postinstall hook, and telemetry preambles. Its 33 SKILL.md files open with telemetry preambles that POST usage events; read them as reference data and don't run the preamble bash blocks. The executables are install.sh, setup and solana-pass.sh at the root plus scripts/bootstrap-agent-setup.sh, scripts/package-skills.sh and two per-skill helpers; install.sh and setup are standalone installers no kit installer runs — both document a curl | bash route, setup git clones sendaifun/solana-new-cli into ~/.claude/skills/superstack and runs itself from there, and install.sh creates ~/.claude/skills and ~/.codex/skills and writes a ~/.superstack config directory. Its root package.json declares postinstall: node dist/cli/index.js init --agent, which would run its CLI's init on any npm install inside the pack; dist/ is not in the pinned checkout and the command ends in `|| true`, so it is a no-op here, but it is a real install lifecycle hook. Eleven files carry a pipe-to-shell line, including its own README, CLAUDE.md and public/setup.sh. child_process, spawn and execSync run throughout cli/, which is the CLI the installers build. One U+200D sits in an emoji sequence in skills/data/ideas/superteam-ideas.json, and the eval and exec hits are detection patterns in its bundled cso skill. No hooks.json, no .claude/settings.json, no .mcp.json, no symlinks, no .gitmodules of its own, no nested .claude/skills/. It ships both AGENTS.md and CLAUDE.md at its root as real files, the nested instruction files an agent working inside the pack directory picks up.",
"tags": ["go-to-market", "design", "marketing", "research", "telemetry"]
},
{
"id": "content-gen-skill",
"name": "ContentGen Skill",
"type": "skill",
"tier": "extension",
"domain": "education",
"path": ".claude/skills/ext/content-gen-skill",
"commit": "ce49a7e82b7120dd49be22c015648f9deeb3b86e",
"description": "Writes educational Solana content in eight forms from one router: course, tutorial, walkthrough, explainer, essay, litepaper, slides and post. A form router picks the pipeline, then every piece runs the same fixed stages: Bloom-tagged backward design before any module list, a Solana prerequisite DAG that fixes lesson order, claims grounded against live sources rather than model memory, draft, an optional writer-style voice pass, a visual-placeholder pass, and validators. Courses also emit a validated per-lesson brief filesystem under content/courses/. Ships twelve Python tools (scaffold, validate, verify blocks and code and challenges, quiz balance, academy export, visual render) and a Superteam Brazil brand asset set.",
"triggers": ["Solana course or curriculum", "tutorial", "explainer or deep-dive", "litepaper", "slide deck text spec", "tweet thread"],
"source": "https://github.com/solanabr/content-gen-skill",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add content-gen-skill", "env": [] },
"license": "MIT",
"maintainer": "solanabr",
"signal": { "stars": 1, "last_commit": "2026-09-04", "reputability": "org" },
"default_installed": false,
"safety": "no executables at this pin, and it writes text and never deploys anything. No hooks.json, no .claude/settings.json, no .mcp.json, no symlinks, no package.json and so no install lifecycle hook, no .gitmodules of its own, no nested .claude/skills/, and no eval, atob, new Function, createRequire or child_process. Its Python tools run locally on the course tree, and the opt-in render step needs WeasyPrint and a rasterizer. One U+200D sits in an emoji sequence in a tutorial exemplar, and three files carry a pipe-to-shell line, all inside worked-example prose (rustup, Anza, a Dockerfile). It ships CLAUDE.md at its root, the nested instruction file an agent working inside the pack directory picks up, the same way it would an AGENTS.md there. Its own .env.example is a stale copy of an older kit env file, including a retired COLOSSEUM_COPILOT_PAT: use the kit's .env.example instead. Largest of the three solanabr content packs on disk (about 7.7 MB, mostly brand PNGs) and ~165 tokens of session listing.",
"tags": ["content", "education", "courses", "writing", "solana"]
},
{
"id": "writer-style-skill",
"name": "Writer-Style Skill",
"type": "skill",
"tier": "extension",
"domain": "ux-writing",
"path": ".claude/skills/ext/writer-style-skill",
"commit": "77372715fb5596cd29cc9fbbcaeb859cc9335b39",
"description": "Writes original long-form and educational prose in a named author's voice, facts first and style second: technical claims are verified before any styling pass, then a two-layer voice pack applies an always-on primary idiolect plus secondary craft borrowed from master writers and routed by the piece's job. Ships the kaue pack (Solana and Web3 education, the default) and the david pack (course lessons, Medium long-form, delivery artifacts), a persona-builder agent for new packs, and validators that strip AI tells. content-gen-skill hands it a brief tagged with dominant_job when both are installed.",
"triggers": ["write in my voice", "write as a named author", "humanize AI prose", "long-form technical writing", "build a voice pack"],
"source": "https://github.com/solanabr/writer-style-skill",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add writer-style-skill", "env": [] },
"license": "MIT",
"maintainer": "solanabr",
"signal": { "stars": 2, "last_commit": "2026-09-04", "reputability": "org" },
"default_installed": false,
"safety": "one executable wired to a postinstall hook, and otherwise prose only with no network or on-chain writes. bin/install.js is the npm standalone installer and package.json declares postinstall: node bin/install.js, so an npm install inside the pack directory would run it — no kit installer does that, and the kit vendors the pack instead. What it does when run: copies the skill, agents, commands and rules into the current project's ./.claude/ with Node core only, and shells out once for `python3 --version` to probe for the validator's interpreter. No hooks.json, no .claude/settings.json, no .mcp.json, no symlinks, no .gitmodules of its own, no nested .claude/skills/, no eval, atob, new Function or createRequire. Two U+FEFF characters appear inside regex literals in tools/validate_voice.py and testbed/check_round.py that strip a BOM — detection, not payload. It ships CLAUDE.md at its root, the nested instruction file an agent working inside the pack directory picks up, the same way it would an AGENTS.md there. Voice fidelity is corpus-bounded: the pack states that its output is approximate and that the secondary profiles are craft rather than impersonation, so fact-check before publishing. The most expensive of the three in session listing (~296 tokens, above solana-dev's ~253).",
"tags": ["writing", "voice", "content", "education", "long-form"]
},
{
"id": "sign-safe",
"name": "Sign-Safe",
"type": "skill",
"tier": "extension",
"domain": "security",
"path": ".claude/skills/ext/sign-safe",
"commit": "e288e9dc9760501399b02e905e1007fa210b9ddf",
"description": "Signing-time safety gate: decodes an opaque base64 transaction or message (legacy and v0 with Address Lookup Tables), classifies instructions against a danger-primitive catalog, screens recipients against a drainer blocklist, and emits a SIGN / HOLD / REJECT verdict plus a machine-readable verdict.json for gating an agent's signing. Offline and deterministic by default. Complements auditor-skill: that audits the program, this one reviews the transaction about to be signed.",
"triggers": ["is this transaction safe to sign", "review a tx before signing", "decode a base64 transaction", "blind signing", "Squads proposal review"],
"source": "https://github.com/lrafasouza/sign-safe-skill",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add sign-safe", "env": [] },
"license": "MIT",
"maintainer": "lrafasouza",
"signal": { "stars": 1, "last_commit": "2026-06-30", "reputability": "individual" },
"default_installed": false,
"safety": "clean at this pin: no eval, atob, createRequire or new Function anywhere; no network in the deterministic core, and its own tests assert no http/axios/undici import leaks into it; spawnSync and execFileSync appear only in test harnesses, with shell false and static argument arrays; no install lifecycle hooks. Egress only through an --rpc endpoint you pass yourself. About 5.8 MB checked out, 93% of it a dev-only mainnet transaction corpus the runtime never reads (upstream issue 29)",
"tags": ["security", "transaction-review", "signing", "drainer", "squads"]
},
{
"id": "counterparty-gate",
"name": "Solana Counterparty Gate",
"type": "skill",
"tier": "extension",
"domain": "security",
"path": ".claude/skills/ext/counterparty-gate",
"commit": "8df0ede96e7a706404b338f390bfa101f0bdeb40",
"vendored": { "skill/solana-dev": "99be3ef2b55587cbf4cab2bd346160eb75e944a0" },
"description": "Operator-reputation gate for a program, oracle, keeper or multisig you did not write: asks whether the deployer behind it is a known serial rug operator before you CPI into it or compose with it. Answers the question a bytecode audit cannot. Backed by SolSentry's hosted mainnet scanner, so its verdicts are only as current as that service.",
"triggers": ["who operates this program", "is this deployer a rug operator", "counterparty risk", "vet a protocol before integrating", "check a multisig operator"],
"source": "https://github.com/solsentry/solana-counterparty-gate",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add counterparty-gate", "env": [] },
"license": "MIT",
"maintainer": "solsentry",
"signal": { "stars": 2, "last_commit": "2026-09-29", "reputability": "individual" },
"default_installed": false,
"safety": "three executables, two of them global installers, and the smallest pack the kit pins at about 144 KB. install.sh and install-custom.sh are for standalone installs and no kit installer runs them: they rm -rf and replace $HOME/.claude/skills/solana-counterparty-gate, and fetch solana-dev-skill from github.com into $HOME/.claude/skills/solana-dev alongside it. tests/smoke.sh is the third, and curls the API below. No eval, atob, createRequire or code-execution primitives; no hooks.json, no .claude/settings.json, no .mcp.json, no symlinks, no package.json and so no install lifecycle hook, no nested .claude/skills/, no bidi or invisible Unicode. It declares one submodule of its own (skill/solana-dev), fetched by no kit installer. It ships both AGENTS.md and CLAUDE.md at its root as real files, the nested instruction files an agent working inside the pack directory picks up. Caution on data flow rather than code: it queries SolSentry's hosted API at api.solsentry.app (read endpoints need no key; override with SOLSENTRY_API_URL), so the counterparty addresses you ask about go to a third party, and an outage degrades it to no answer rather than a wrong one.",
"tags": ["security", "counterparty", "operator-reputation", "due-diligence", "api-backed"]
},
{
"id": "community-moderation",
"name": "Community Moderation (Foka AI)",
"type": "skill",
"tier": "extension",
"domain": "go-to-market",
"path": ".claude/skills/ext/community-moderation",
"commit": "e10790ea57cba41626704560a9355b2761d77816",
"description": "Moderate Solana community groups on Telegram and Discord and run member support: detect and act on spam, drainer links, raids, admin impersonation and mass pings; member trust states, roles and reputation; support triage, tagging and ticketing; Combot and MEE6 parity features. Evasion-resistant detection across ten languages, with confusable, invisible-character and bidi normalization.",
"triggers": ["moderate a Telegram or Discord group", "detect scam and drainer links", "community raid response", "admin impersonation", "support ticket triage"],
"source": "https://github.com/soufoka/community-moderation-skill",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add community-moderation", "env": [] },
"license": "MIT",
"maintainer": "soufoka",
"signal": { "stars": 2, "last_commit": "2026-06-30", "reputability": "individual" },
"default_installed": false,
"safety": "clean at this pin: no eval, atob, createRequire or new Function; no install lifecycle hooks; one runtime dependency (confusables); bin/install.js copies files with Node core only and no network. The invisible-Unicode and bidi characters in examples/normalize.ts and examples/content-filters.ts are detection character classes, not payload — verified by reading them. Its Telegram, Discord and WhatsApp examples read bot tokens from env and talk only to those platforms' own APIs; running them is the operator's choice, not the skill's",
"tags": ["community", "moderation", "telegram", "discord", "support"]
},
{
"id": "frontend-slides",
"name": "Frontend Slides",
"type": "skill",
"tier": "extension",
"domain": "frontend-design",
"path": ".claude/skills/ext/frontend-slides",
"commit": "9906a34d640d2111f724544cbc50f7f130569ae1",
"description": "Animation-rich HTML slide decks, and the house format for decks, graphics and marketing material. 36 designed templates with per-template design.md and preview.md, a selection index that maps a deck's intent to a template, a shared deck-stage.js runtime, and scripts to render a deck to PDF or publish it. HTML in, no PowerPoint: the kit's content-gen skill hands it the content.",
"triggers": ["slide deck or presentation in HTML", "pitch deck or demo-day slides", "marketing graphic or social image", "export a deck to PDF", "publish a deck to a URL"],
"source": "https://github.com/zarazhangrui/frontend-slides",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add frontend-slides", "env": [] },
"license": "MIT",
"maintainer": "zarazhangrui",
"signal": { "stars": 30126, "last_commit": "2026-06-23", "reputability": "individual" },
"default_installed": false,
"safety": "Scanned at 9906a34 (2026-06-23): no lines over 500 characters, no eval, atob, new Function or createRequire, no bidi or invisible Unicode, no IP literals, no package.json and so no install lifecycle hook, no hooks.json, no .mcp.json and no .claude/settings.json; all 291 blobs across its 31 commits hold nothing the checkout does not, and nothing has ever been deleted from it. Two executables sit inside the skill directory, each also duplicated at the repo root, and neither runs unless you ask for an export or a deploy. scripts/export-pdf.sh writes a Node script into a temp dir, runs npm install playwright and npx playwright install chromium there, renders the deck headless and rm -rf's the temp dir after. scripts/deploy.sh publishes a deck through the Vercel CLI and will npm install -g vercel if neither a global vercel nor npx can provide it, so a deploy both installs a global package and puts your slides on a public URL under your own Vercel account. Decks fetch webfonts at render time from fonts.googleapis.com, fonts.gstatic.com, api.fontshare.com and chinese-fonts-cdn.deno.dev, and templates reference raw.githubusercontent.com and cdn.jsdelivr.net assets, so a deck is not offline-clean. Personal-account repo, vetted in the provenance table.",
"tags": ["slides", "presentation", "html", "animation", "frontend", "marketing"]
},
{
"id": "superseo",
"name": "SuperSEO Skills",
"type": "skill",
"tier": "extension",
"domain": "marketing",
"path": ".claude/skills/ext/superseo",
"commit": "9cf22cc065a294bf0077ea6af6b4ea0eb7f89cdc",
"description": "Eleven SEO skills for the pages a Solana project actually ships: page audits, content briefs, article writing with an anti-AI-slop ruleset, E-E-A-T and YMYL scoring, semantic gap analysis, topic clusters, featured-snippet optimization, link building and expert interviews. Bundles 23 content-type templates, 16 writing-technique modules and 9 link-building playbooks. Drives your own search and fetch tools and names no paid data source.",
"triggers": ["SEO audit of a landing page", "write or improve a blog post or docs page", "keyword research or search intent", "topic clusters and internal linking", "why does this page not rank"],
"source": "https://github.com/inhouseseo/superseo-skills",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add superseo", "env": [] },
"license": "Apache-2.0",
"maintainer": "inhouseseo",
"signal": { "stars": 343, "last_commit": "2026-09-03", "reputability": "individual" },
"default_installed": false,
"safety": "Scanned at 9cf22cc (2026-09-03): zero executables and zero script files of any kind — no .sh, .py, .js, .mjs, .cjs or .ts anywhere. The pack is markdown plus one demo GIF and one asciinema cast. No package.json at all and so no install lifecycle hook; no hooks.json, no .mcp.json, no .claude/settings.json; no eval, atob, new Function or createRequire; no child_process; no lines over 500 characters; no bidi or invisible Unicode; no IP literals; and no env-var or credential read outside its own GitHub release workflow, which only zips the skills folder onto a release. Its 415 blobs across 54 commits add only a deleted demo/demo-runner.sh and a research/ tree since folded into skills/*/references/, neither of which a pin checks out. Verbatim 202-line Apache-2.0 plus a NOTICE naming InhouseSEO, and a SECURITY.md whose claims match what the scan found. It works through your own WebSearch and WebFetch tools: the SEO vendors named in its text (ahrefs.com, semrush.com, backlinko.com, searchengineland.com) are prose citations, not endpoints it calls, so no paid data source is required and nothing is sent anywhere you did not already send it. Provenance is the thin part, not the code: the author account is six months old with this as its only repository, which is why it has a row in the provenance table.",
"tags": ["seo", "content", "marketing", "copywriting", "eeat", "link-building"]
},
{
"id": "expo",
"name": "Expo Skills",
"type": "skill",
"tier": "extension",
"domain": "mobile",
"path": ".claude/skills/ext/expo",
"commit": "a9e5bd3fa00b4e61f632e43f85da73aee1942582",
"vendored": { "eval-harness": "da0eba34bf02b6dc57a6eca611c4bdab73b246c5" },
"description": "Expo's own 23 skills for the mobile half of a Solana project, and the part solana-mobile does not cover: EAS Build, Update and Workflows CI, app-store submission, OTA updates and rollouts, dev clients, config plugins and native modules, expo-router, expo-ui and the design system, brownfield embedding into an existing native app, SDK upgrades, and EAS Hosting and Observe. Pair it with solana-mobile for Mobile Wallet Adapter and Seeker; this one is the build, release and CI side.",
"triggers": ["EAS build or submit to the app stores", "over-the-air update or rollout", "Expo config plugin or native module", "expo-router navigation or expo-ui", "upgrade the Expo SDK"],
"source": "https://github.com/expo/skills",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add expo", "env": [] },
"license": "MIT",
"maintainer": "expo",
"signal": { "stars": 2655, "last_commit": "2026-10-03", "reputability": "official" },
"default_installed": false,
"safety": "it ships its own .claude/skills/ directory, which Claude Code auto-discovers, and ten executables of which eight live inside it. Claude Code lists any directory named .claude/skills at any depth, so in a project that installs this pack expo-skill-eval (48 KB of SKILL.md, a third-party skill that boots iOS and Android simulators and runs bunx expo) is registered and invokable with no further action, and its description joins every session and every subagent listing. install.sh and skills.sh strip a pack's own .claude/ from the copy that lands in a project (.claude/bin/_pack_strip.sh), so that skill no longer reaches an installed project, and tests/test_pack_load_surfaces.sh asserts it. That is a property of where this pack puts its skills, not of ext/, and it is the one pinned pack that does it — Codex and opencode discover nested SKILL.md files the same way. The other two executables are scripts/ci.sh and ci.py, its CI, where EVAL_HARNESS_ACCESS_TOKEN clones a private harness. Scanned at 13ad8e0 (2026-10-03): no lines over 500 characters, no bidi or invisible Unicode, no install lifecycle hook in any of its 4 package.json files, no .claude/settings.json, and its 3,517 blobs across 268 commits hold no keys, no binaries and no deleted hook config. It ships one symlink, CLAUDE.md -> AGENTS.md at the pack root, resolving to the real AGENTS.md beside it, and declares one submodule of its own (eval-harness) that no kit installer fetches. The eval and atob grep hits are inert: one atob in a React Native sample in expo-native-ui/references/media.md, and the rest are Swift and Kotlin Function and AsyncFunction declarations from the Expo Modules API in documentation. It ships two more things that fire only if you install it as a Claude Code plugin, never at this ext/ pin: plugins/expo/hooks/hooks.json runs node skills/expo-skill-feedback/scripts/skill-event.cjs on PostToolUse with matcher Skill and on UserPromptExpansion, and .mcp.json declares a remote HTTP MCP server at https://mcp.expo.dev/mcp, which names no npm package and so carries no dependency-confusion surface, though enabling it does send your Expo and EAS queries to Expo. Telemetry verified by reading telemetry_common.cjs and skill-event.cjs rather than taking the README's word: it is opt-in with DEFAULT_ON set to false and sends nothing until you create the marker file at ~/.expo-skills/opt-in or set EXPO_SKILLS_TELEMETRY=1; DO_NOT_TRACK=1 forces it off with the highest precedence; CI detection forces it off too, covering CI plus GITHUB_ACTIONS, GITLAB_CI, CIRCLECI, TRAVIS, BUILDKITE, JENKINS_URL, TEAMCITY_VERSION and TF_BUILD. When it is on, the event carries only the skill name, initiator, OS, arch and a truncated SHA-256 of a random per-install UUID stored 0600 at ~/.expo-skills/installation-id, POSTed to us.i.posthog.com with a 3-second timeout and $process_person_profile false. No prompt text, no file paths, no project or repository names. The phc_ PostHog key committed in that file is a write-only public ingestion key, documented as such in place.",
"tags": ["mobile", "expo", "react-native", "eas", "ci", "app-store", "ota"]
},
{
"id": "duckdb",
"name": "DuckDB Skills",
"type": "skill",
"tier": "extension",
"domain": "data",
"path": ".claude/skills/ext/duckdb",
"commit": "7feda8e01e22bc0886c86123f3884947e36d8c69",
"description": "Query and report on data without standing up a database: nine skills over reading any data file (Parquet, CSV, JSON, Avro, Excel, Iceberg, Delta), running SQL, attaching DuckDB and DuckLake databases, converting between formats, exploring S3, R2, GCS and Azure buckets, geospatial work, and searching DuckDB's docs. The right tool for an indexer's Parquet exports, an airdrop snapshot or a one-off analysis over transaction dumps, where Postgres would be overkill.",
"triggers": ["analyze a CSV Parquet or JSON dump", "SQL over local or S3 files", "convert between data formats", "report or aggregate over exported onchain data", "attach and query a DuckDB or DuckLake database"],
"source": "https://github.com/duckdb/duckdb-skills",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add duckdb", "env": [] },
"license": "MIT",
"maintainer": "duckdb",
"signal": { "stars": 598, "last_commit": "2026-04-23", "reputability": "official" },
"default_installed": false,
"safety": "Scanned at 7feda8e (2026-04-23): 17 files, no lines over 500 characters, no eval, atob, new Function or createRequire, no child_process, no bidi or invisible Unicode, no IP literals, no package.json and so no install lifecycle hook, no hooks.json, no .mcp.json and no .claude/settings.json; its 291 blobs across 61 commits add nothing the checkout lacks and nothing has been deleted. One executable, skills/install-duckdb/eval.sh, a developer eval harness that shells out to claude --print and duckdb and which no SKILL.md references. Two things to know before using it. First, install-duckdb/SKILL.md tells you to install DuckDB with curl -fsSL https://install.duckdb.org | sh: that is a pipe-to-shell it asks you to run, not one it runs itself, and the DuckDB CLI is also in Homebrew and the usual package managers if you would rather not. Second, the read-memories skill searches your past Claude session logs on disk, which is useful and also means it reads transcripts of other work. Its S3, GCS and Azure skills issue DuckDB CREATE SECRET statements with PROVIDER credential_chain, which hands authentication to the cloud SDK's own credential chain rather than reading a credential itself, but a query against a remote bucket does send your data request off the machine. MIT, Stichting DuckDB Foundation.",
"tags": ["data", "sql", "duckdb", "parquet", "analytics", "reporting", "csv"]
},
{
"id": "crypto-legal",
"name": "Crypto Legal Skill",
"type": "skill",
"tier": "extension",
"domain": "compliance",
"path": ".claude/skills/ext/crypto-legal",
"commit": "0f0de7ccbf39213c20bc34fa76791628ad43b838",
"description": "The legal and compliance questions a token launch raises, with statutory citations rather than vibes: is this mint a security, is the airdrop lawful, what does the Terms of Service need, is a money-transmitter licence or BitLicense in scope. Seven domain references (securities law, AML and KYC, sanctions, privacy and data protection, tax, tokenomics legality, Solana-specific) across three jurisdictions (US, EU under MiCA, Brazil under Lei 14.478 and the BCB resolutions), plus triage and launch-checklist workflows and document checklists. Informational only, never a substitute for counsel.",
"triggers": ["is this token a security", "is my airdrop legal", "review a Terms of Service or privacy policy", "do I need a money transmitter licence or BitLicense", "sanctions or KYC obligations for a launch"],
"source": "https://github.com/solanabr/crypto-legal-skill",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add crypto-legal", "env": [] },
"license": "MIT",
"maintainer": "solanabr",
"signal": { "stars": 2, "last_commit": "2026-07-10", "reputability": "first-party" },
"default_installed": false,
"safety": "Scanned at 0f0de7c (2026-07-10): 42 files, no lines over 500 characters, no eval, atob, new Function or createRequire, no child_process, no bidi or invisible Unicode, no IP literals, no package.json and so no install lifecycle hook, no hooks.json, no .mcp.json and no .claude/settings.json; 94 blobs across 4 commits with nothing deleted, and install.sh is the only script path the repo has ever had. The kit never runs that installer: it exists for people installing the skill standalone, and it writes to ~/.claude/skills/crypto-legal, ~/.claude/commands/crypto-legal/ and, when the codex CLI is on PATH, ~/.codex/skills/crypto-legal; it is interactive and can move an existing install aside to a .backup.<timestamp> folder; it makes no network call. Checked specifically because a different repo of the same name was rejected for it: this one does not write an unscoped rule into ~/.claude/. Its two rules/*.md files use globs: frontmatter, which Claude Code ignores, but install.sh copies them inside the skill folder and never into ~/.claude/rules/, and validate.sh's unscoped-rule check only scans .claude/rules, so they can neither load every session nor fail a gate from this pin. Pinned whole for that reason rather than skill/ alone. What matters more than any of that is what the content is: informational only, not legal advice, no attorney-client relationship, and no substitute for licensed counsel in the relevant jurisdiction. Its statutory review is calendar-pinned at 2026-06 and goes stale as law moves, and it covers the US, the EU and Brazil only — treat an answer about anywhere else as absent, not as permissive. First-party solanabr repo, the same organisation that publishes this kit, so it needs no provenance-table row.",
"tags": ["legal", "compliance", "securities", "aml", "kyc", "privacy", "tax", "mica"]
},
{
"id": "gsap-skills",
"name": "GSAP Skills",
"type": "skill",
"tier": "extension",
"domain": "frontend-design",
"path": ".claude/skills/ext/gsap-skills",
"commit": "aed9cfd3277740755f6bfc1155c7aa645403b760",
"description": "GreenSock's own pack for GSAP, in eight skills: the core tween and property API, timelines and sequencing, ScrollTrigger, the plugin set, the React integration (useGSAP), framework wiring for Vue and Nuxt, the utility methods, and a performance lane. The animation craft layer for a dApp's landing page or marketing site — solana-new carries page-load references and the kit has no other GSAP coverage.",
"triggers": ["GSAP", "ScrollTrigger", "timeline animation", "scroll-driven animation", "useGSAP", "landing page motion"],
"source": "https://github.com/greensock/gsap-skills",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add gsap-skills", "env": [] },
"license": "MIT",
"maintainer": "greensock",
"signal": { "stars": 15939, "last_commit": "2026-04-21", "reputability": "official" },
"default_installed": false,
"safety": "Scanned at aed9cfd (2026-04-21): 264 KB, eight skill folders, and clean on every checklist item — zero executables, no .mcp.json, no hooks.json, no .claude/settings.json, no eval, atob, new Function or createRequire, no child_process, no bidi or invisible Unicode, no lines over 500 characters, and no environment variable or credential read anywhere. Verbatim MIT at the repo root. Three package.json files exist, all under examples/; examples/nuxt declares postinstall: nuxt prepare, which runs only if someone runs npm install inside that example directory, and no kit installer does. Its one caveat is the reason it is pinned as a submodule rather than fetched folder-by-folder: it ships two symlinks at its root, CLAUDE.md -> AGENTS.md and GEMINI.md -> AGENTS.md, and both resolve to the real AGENTS.md beside them. That is the shape ext/vercel has shipped since it was pinned, and the same end state helius, qedgen, solana-game, solana-new and mongodb already produce with plain files, so it is a pre-existing condition of the catalogue rather than something this pack introduces — see the copy-mechanics table in docs/skill-packs.md and do not re-litigate it per pack. An ext/ submodule also costs nothing in standing context, where fetching its eight folders top-level would have added about 674 tokens to every session and every subagent. Last moved 2026-04-21, around five months before this pin, so check its API against the current GSAP release.",
"tags": ["gsap", "animation", "scrolltrigger", "motion", "frontend", "official"]
},
{
"id": "dune",
"name": "Dune Analytics Skill",
"type": "skill",
"tier": "extension",
"domain": "data-analytics",
"path": ".claude/skills/ext/dune",
"commit": "cf5c7fae42b03ed653bd49ae44d775f5c803ce72",
"description": "Dune's own skill for writing and debugging DuneSQL: query authoring against Dune's Solana and EVM tables, the Trino dialect's departures from Postgres, materialised views and query performance. Complements an indexer rather than replacing one — Dune answers historical and cross-chain questions over data someone else already decoded.",
"triggers": ["Dune query", "DuneSQL", "onchain analytics", "dashboard query", "historical chain data"],
"source": "https://github.com/duneanalytics/skills",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add dune", "env": [] },
"license": "MIT",
"maintainer": "duneanalytics",
"signal": { "stars": 52, "last_commit": "2026-09-14", "reputability": "official" },
"default_installed": false,
"safety": "Scanned at cf5c7fa (2026-09-14): the smallest pack pinned here at 116 KB, one skill folder, markdown and SQL only. No executables, no package.json and so no install lifecycle hook, no hooks.json, no .mcp.json, no symlinks, no bidi or invisible Unicode, no lines over 500 characters, no eval, atob, new Function or createRequire, no child_process. No credential read: queries run in Dune's own web app or against its API with a key you supply yourself, and the pack names no environment variable.",
"tags": ["dune", "sql", "analytics", "data", "official"]
},
{
"id": "circle",
"name": "Circle Skills",
"type": "skill",
"tier": "extension",
"domain": "payments",
"path": ".claude/skills/ext/circle",
"commit": "1f23655cbc360f7783e6881e026dbd4568c40f1d",
"description": "Circle's own pack for USDC and stablecoin payments: 18 skills over developer- and user-controlled wallets, modular and agent wallets, CCTP bridging, Gateway, Arc, the Smart Contract Platform, token swaps, balance unification and agent-to-agent payments. The fiat-rails counterpart to the kit's on-chain DeFi lane, and the route to USDC settlement that does not go through a DEX.",
"triggers": ["USDC payments", "stablecoin settlement", "CCTP bridge", "agent wallet", "accept payments", "Circle API"],
"source": "https://github.com/circlefin/skills",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add circle", "env": ["CIRCLE_API_KEY", "CIRCLE_ENTITY_SECRET"] },
"license": "Apache-2.0",
"maintainer": "circlefin",
"signal": { "stars": 155, "last_commit": "2026-09-16", "reputability": "official" },
"default_installed": false,
"safety": "caution — this pack moves real money, and that is its purpose rather than a defect. Scanned at 58ab864 (2026-09-16): 1.6 MB, 18 skills, no symlinks, no hooks.json, no .claude/settings.json, no package.json and so no install lifecycle hook, no bidi or invisible Unicode, no eval, atob, new Function or createRequire, no child_process. Two executables, both under skills/recover-eco-funds/scripts/: eco-self-refund.mjs and eoa-to-sca.mjs. Their default mode is read-only and simulates the refund calldata; signing is opt-in behind explicit flags, and they write an evidence directory. Eleven of the 18 skills describe fund movement — pay-via-agent-wallet, fund-agent-wallet, accept-agent-payments, swap-tokens, bridge-stablecoin, unify-balance and recover-eco-funds among them — and are prose instructions for Circle's API and CLI, not executables. The credential surface is the widest of any pack pinned here: pack text and scripts name CIRCLE_API_KEY, CIRCLE_ENTITY_SECRET, ENTITY_SECRET, DELEGATE_PRIVATE_KEY, EVM_PRIVATE_KEY, PRIVATE_KEY, SOLANA_PRIVATE_KEY and SOLANA_PRIVATE_KEYPAIR. Its one MCP entry (plugins/circle/.mcp.json) is remote and points at https://api.circle.com/v1/codegen/mcp; it names no npm or PyPI package. Treat every write path as production until you have pointed it at a testnet, and read agent-wallet-policy before granting an agent wallet any allowance.",
"tags": ["usdc", "stablecoin", "payments", "wallets", "cctp", "official", "moves-funds"]
},
{
"id": "knowledge-work",
"name": "Anthropic Knowledge Work Plugins",
"type": "skill",
"tier": "extension",
"domain": "dev-workflow",
"path": ".claude/skills/ext/knowledge-work",
"commit": "2574e29377c7173d0cecffef29a6ecbd7f8ccb5d",
"description": "Anthropic's knowledge-work plugin collection: 252 skills across 18 role-shaped plugins — engineering, product management, design, data, finance, legal, sales, marketing, operations, human resources, customer support, enterprise search, bio-research, small-business and a partner-built set. The business-operations layer around a Solana project (fundraising memos, contracts, support, go-to-market) rather than anything on-chain, which is why it is an extension and not core.",
"triggers": ["business operations", "go-to-market docs", "enterprise search", "sales or support workflow", "role-specific knowledge work"],
"source": "https://github.com/anthropics/knowledge-work-plugins",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add knowledge-work", "env": [] },
"license": "Apache-2.0",
"maintainer": "anthropics",
"signal": { "stars": 26085, "last_commit": "2026-10-01", "reputability": "official" },
"default_installed": false,
"safety": "caution for its connector surface, which is the largest of any pack pinned here. Scanned at 8444efc (2026-10-01): 11 MB, 252 skills, no symlinks, no hooks.json, no .claude/settings.json, no package.json and so no install lifecycle hook, no bidi or invisible Unicode, no eval, atob, new Function or createRequire, no child_process. 21 .mcp.json files declare 186 MCP server entries: 185 remote HTTPS endpoints across roughly 90 distinct third-party hosts (Slack, Notion, Atlassian, HubSpot, Salesforce, Google, Microsoft 365, Box, Stripe, PayPal, Shopify, Zoho, Zoom and many more), and one stdio entry, pdf-viewer's npx -y @modelcontextprotocol/server-pdf --stdio, which exists on the npm registry (latest 2.0.3) and is the only npm or PyPI package any of these files names. 30 of the 186 entries carry an empty \"url\": \"\" and are placeholders that cannot connect until someone fills them in — benchling, snowflake and databricks, and the google calendar, gmail and google drive rows in 12 of the role plugins. A .mcp.json in a pinned pack does not take effect from the pin: nothing in ext/ is read as project MCP config, so these are a catalogue of connectors you would have to add deliberately. Each one you do add authenticates as you and sends the data you ask about to that vendor. A handful of skills ship Python under scripts/ (bio-research's single-cell RNA QC among them) which run locally on files you point at.",
"tags": ["knowledge-work", "business", "mcp-connectors", "productivity", "official"]
},
{
"id": "mongodb",
"name": "MongoDB Agent Skills",
"type": "skill",
"tier": "extension",
"domain": "data-analytics",
"path": ".claude/skills/ext/mongodb",
"commit": "155bb9c4b42a9a3e019f47be0669de0e4547d1ea",
"description": "MongoDB's own pack: schema design, query optimisation, natural-language querying, Atlas Search and vector/AI search, Atlas Stream Processing, connection handling, queryable-encryption size estimation and MCP setup. For a Solana project this is the off-chain store under an indexer — the document-database counterpart to the duckdb and supabase extensions.",
"triggers": ["MongoDB", "Atlas Search", "schema design", "aggregation pipeline", "vector search", "indexer datastore"],
"source": "https://github.com/mongodb/agent-skills",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add mongodb", "env": [] },
"license": "Apache-2.0",
"maintainer": "mongodb",
"signal": { "stars": 188, "last_commit": "2026-10-02", "reputability": "official" },
"default_installed": false,
"safety": "Scanned at 18b014e (2026-10-02): 3.2 MB, 24 SKILL.md files (the plugins/ tree duplicates the top-level skills/ tree), no symlinks, no hooks.json, no .claude/settings.json, no bidi or invisible Unicode, no lines over 500 characters. Three executables, all repo maintenance rather than anything a user runs: tools/validate-skills.sh and .github/scripts/{validate-skills,qa-resolve-skills}.sh. Four files reference child_process or an eval harness — tools/plugin-version.ts, tools/sync-plugin-skills.ts, testing/lint-item-echo.mjs and testing/mongodb-schema-design/run-trigger-eval.ts — all in the repo's own CI and test tooling, none reachable from a SKILL.md. Its package.json declares no install lifecycle hook. One MCP entry (plugins/mongodb/mcp.json) is stdio: npx -y mongodb-mcp-server@<4 (re-checked at pin 155bb9c4 on 2026-11-10; #326 widened it from @<3), which exists on the npm registry (latest 3.3.0, so that range resolves to the latest 3.x); it would connect to whichever cluster its connection string names, and a pinned .mcp.json does not take effect until you add it yourself. It ships GEMINI.md at its root, a nested instruction file an agent working inside the pack directory could pick up — the same shape five already-pinned packs have.",
"tags": ["mongodb", "database", "atlas", "search", "vector", "official"]
},
{
"id": "elevenlabs",
"name": "ElevenLabs Skills",
"type": "skill",
"tier": "extension",
"domain": "go-to-market",
"path": ".claude/skills/ext/elevenlabs",
"commit": "32f48f6ebf4a97e963a824ade7d5cb14beacf774",
"description": "ElevenLabs' own pack: text-to-speech, speech-to-text, voice changer, voice isolator, dubbing, sound effects, music and conversational agents. The audio lane for a launch — demo-video narration, localisation and trailer sound — next to the kit's content-gen and marketing-video routes.",
"triggers": ["text to speech", "voiceover", "demo video narration", "dubbing", "sound effects", "transcription"],
"source": "https://github.com/elevenlabs/skills",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add elevenlabs", "env": ["ELEVENLABS_API_KEY"] },
"license": "MIT",
"maintainer": "elevenlabs",
"signal": { "stars": 471, "last_commit": "2026-10-02", "reputability": "official" },
"default_installed": false,
"safety": "Scanned at 81f1eaf (2026-10-02): 868 KB, 10 skills, no symlinks, no hooks.json, no .claude/settings.json, no package.json and so no install lifecycle hook, no bidi or invisible Unicode, no eval, atob, new Function or createRequire, no child_process. One executable, evals/run_all.py, the repo's own eval runner. Its setup-api-key skill walks you through ELEVENLABS_API_KEY; every generation call is paid and billed to that key, so a long dubbing or music job costs real money. Its one MCP entry (mcp.json) is remote and points at https://api.us.elevenlabs.io/v1/mcp — a US endpoint, which is also where the audio you send is processed; it names no npm or PyPI package.",
"tags": ["audio", "tts", "voice", "dubbing", "marketing", "official"]
},
{
"id": "light-protocol",
"name": "Light Protocol ZK Compression Skill",
"type": "skill",
"tier": "extension",
"domain": "solana-core",
"path": ".claude/skills/ext/light-protocol",
"commit": "0f52958edd1872d43c9d94fd58f558abb250789a",
"description": "Light Protocol's own pack for ZK compression on Solana: compressed PDAs, compressed tokens, the ZK primitives behind them, and a testing lane. The route to state that costs a fraction of rent, which no other pinned pack covers — the compression counterpart to the kit's ordinary account work.",
"triggers": ["ZK compression", "compressed token", "compressed PDA", "state cost", "Light Protocol"],
"source": "https://github.com/Lightprotocol/skills",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add light-protocol", "env": [] },
"license": "MIT",
"maintainer": "Lightprotocol",
"signal": { "stars": 7, "last_commit": "2026-06-08", "reputability": "official" },
"default_installed": false,
"safety": "Scanned at 0f52958 (2026-06-08): 284 KB, 5 skills, markdown only — no executables, no package.json and so no install lifecycle hook, no hooks.json, no .claude/settings.json, no symlinks, no bidi or invisible Unicode, no eval, atob, new Function or createRequire, no child_process. The thinnest signal of the official packs pinned here (7 stars) and the stalest (last commit 2026-06-08, around four months before this pin), so check its SDK versions against the current crate before following them. Its .mcp.json declares two entries: zkcompression, stdio npx -y mcp-remote@latest https://www.zkcompression.com/mcp — mcp-remote exists on the npm registry (latest 0.14.3) but is requested at @latest rather than pinned, so adding that server would install whatever version npm resolves at the time; and deepwiki, which uses a serverUrl key rather than url and so is not a well-formed entry for current Claude Code. A pinned .mcp.json does not take effect until you add it yourself. It ships CLAUDE.md at its root — the nested instruction file an agent working inside the pack directory picks up, the same way it would an AGENTS.md there.",
"tags": ["zk", "compression", "solana", "compressed-token", "official"]
},
{
"id": "gmgn",
"name": "GMGN Trading Skills",
"type": "skill",
"tier": "extension",
"domain": "solana-defi",
"path": ".claude/skills/ext/gmgn",
"commit": "6a04e1f00b5463cfc965e41003ca0c7c2db5b69f",
"description": "GMGN's OpenAPI pack: 17 skills over token research and due diligence, holder and wallet analysis, smart-money tracking, dev scoring, heat ranks, narrative and K-line pattern reads, portfolio views — and swap execution. Covers the memecoin research surface no other pinned pack reaches; it also places trades, which is the reason to read its safety note before installing.",
"triggers": ["memecoin research", "token due diligence", "holder analysis", "smart money wallet", "swap a token", "limit order"],
"source": "https://github.com/GMGNAI/gmgn-skills",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add gmgn", "env": ["GMGN_API_KEY", "GMGN_PRIVATE_KEY"] },
"license": "MIT",
"maintainer": "GMGNAI",
"signal": { "stars": 595, "last_commit": "2026-09-28", "reputability": "individual" },
"default_installed": false,
"safety": "caution — it executes trades. Two of its 17 skills submit orders: gmgn-swap, whose own description is tagged [FINANCIAL EXECUTION], buys and sells on Solana, BSC, Base and Ethereum and covers single swaps, multi-wallet batch trading, limit orders, stop loss, take profit and trailing orders; gmgn-token-buy resolves a token name to a contract, gates volume, depth and security, sizes slippage and gas, then hands confirmed parameters to gmgn-swap to execute. The skill text states that it requires explicit user confirmation and that it is the only skill holding a key, and src/confirm.ts is a confirmation gate; treat that as the author's design, not a guarantee, and note the pack's own warning that gmgn-swap resolves no token names — only contract addresses — so a same-name copycat is a loss it cannot catch. GMGN_PRIVATE_KEY is a PEM request-signing key for GMGN's API, not a chain wallet key, and the repo's .env.example says so and tells you to chmod 600 it; GMGN_API_KEY is the account credential. Scanned at 4575ef5 (2026-09-28): 3.7 MB, 17 skills, no executables, no symlinks, no hooks.json, no .claude/settings.json, no .mcp.json, no .gitmodules of its own, no nested .claude/skills/. Its package.json declares no install lifecycle hook (only build, dev and prepublishOnly); the CLI it drives, gmgn-cli, is on the npm registry at 1.6.6 and depends on commander, dotenv, undici and socks — that last one means it can route its API calls through a SOCKS proxy if configured. Eleven of its SKILL.md files carry the same 401/403 troubleshooting step, which asks you to curl ipv6.icanhazip.com to find out whether your egress is IPv6; that is a third-party IP-echo service rather than a GMGN host, and it is a command for you to run. Three occurrences of U+200D zero-width joiner appear in each of skills/gmgn-wallet-score/SKILL.md and skills/gmgn-holder-analysis/analyze.py, consistent with emoji sequences in its bilingual text; src/sanitize.ts holds a character class of bidirectional and invisible code points, which is a stripping regex rather than payload, verified by reading it. No bidirectional override characters anywhere else. The repository is a personal account rather than an organisation, and it ships CLAUDE.md at its root — the nested instruction file an agent working inside the pack directory picks up, the same way it would an AGENTS.md there.",
"tags": ["trading", "memecoin", "research", "swap", "solana", "moves-funds"]
},
{
"id": "redis",
"name": "Redis Agent Skills",
"type": "skill",
"tier": "extension",
"domain": "infra",
"path": ".claude/skills/ext/redis",
"commit": "baac7d66798587a19b9aa2c644f83c1db11762a2",
"description": "Redis' own pack: connection handling and pooling, core data structures, clustering, Redis Search, semantic caching for LLM work, security hardening and observability. For a Solana project this is the cache and queue under an RPC fan-out or indexer — the hot-path counterpart to the mongodb and supabase extensions.",
"triggers": ["Redis", "caching", "semantic cache", "Redis Search", "connection pool", "rate limiting store"],
"source": "https://github.com/redis/agent-skills",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add redis", "env": [] },
"license": "MIT",
"maintainer": "redis",
"signal": { "stars": 164, "last_commit": "2026-09-08", "reputability": "official" },
"default_installed": false,
"safety": "Scanned at a84871d (2026-09-08): 1.6 MB, 16 SKILL.md files (the packages/ tree duplicates the top-level skills/ tree), no hooks.json, no .claude/settings.json, no .mcp.json, no bidi or invisible Unicode. Its package.json declares prepare: husky, a lifecycle hook that would run only if someone ran npm install inside the pack directory; no installer the kit ships does that. Three files reference child_process or an eval harness — scripts/validate-skill-structure.mjs and two files under packages/redis-development-build/src/eval/ — all repo CI and benchmark tooling, none reachable from a SKILL.md. It ships one symlink, CLAUDE.md -> AGENTS.md at the pack root, which resolves to the real AGENTS.md beside it. install.sh and skills.sh copy with cp -R and keep it a link; update.sh's frozen copy loop uses cp -r with a plain directory argument, which on BSD/macOS dereferences it into a real CLAUDE.md file (exit 0, no breakage — unlike a dangling link, which would abort the copy). Either way the result is a nested instruction file inside the pack, the same shape helius, qedgen, solana-game, solana-new and vercel already ship.",
"tags": ["redis", "cache", "infra", "observability", "official"]
},
{
"id": "resend",
"name": "Resend Skills",
"type": "skill",
"tier": "extension",
"domain": "infra",
"path": ".claude/skills/ext/resend",
"commit": "b2ae7b1798b70134faf679f35fc05b4331462988",
"description": "Resend's own pack: the Resend CLI, React Email components, deliverability and email-design practice, and an agent email inbox. The transactional-email lane a dApp needs for alerts, receipts and waitlists, which no other pinned pack covers.",
"triggers": ["send email", "transactional email", "React Email", "email template", "deliverability", "waitlist"],
"source": "https://github.com/resend/resend-skills",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add resend", "env": ["RESEND_API_KEY"] },
"license": "MIT",
"maintainer": "resend",
"signal": { "stars": 194, "last_commit": "2026-09-29", "reputability": "official" },
"default_installed": false,
"safety": "Scanned at edbfece (2026-09-29): 656 KB, 5 skills, no executables, no hooks.json, no .claude/settings.json, no package.json lifecycle hook, no bidi or invisible Unicode, no eval, atob, new Function or createRequire, no child_process. Its one MCP entry (mcp.json) is remote and points at https://mcp.resend.com/mcp; it names no npm or PyPI package. RESEND_API_KEY sends mail as your domain, so a mistake reaches real inboxes under your own sending reputation — its agent-email-inbox skill in particular has an agent read and reply to mail. It ships one symlink, CLAUDE.md -> AGENTS.md at the pack root, resolving to the real AGENTS.md beside it; see the redis entry for exactly what each copy path does with it.",
"tags": ["email", "transactional", "react-email", "infra", "official"]
},
{
"id": "pulumi",
"name": "Pulumi Agent Skills",
"type": "skill",
"tier": "extension",
"domain": "infra",
"path": ".claude/skills/ext/pulumi",
"commit": "82f222a529a499e8005a820000148cc5e28c934c",
"description": "Pulumi's own pack: 17 skills over the Automation API, components and package usage, ESC for secrets and config, the Context API, debugging a failed operation, ARM and discovered-stack migration, provider upgrades and upstream patching. Infrastructure-as-code in a real language for the servers under an indexer or RPC fleet, where the kit's own devops lane stops at CI and Cloudflare.",
"triggers": ["Pulumi", "infrastructure as code", "Automation API", "Pulumi ESC", "stack migration", "provider upgrade"],
"source": "https://github.com/pulumi/agent-skills",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add pulumi", "env": [] },
"license": "Apache-2.0",
"maintainer": "pulumi",
"signal": { "stars": 70, "last_commit": "2026-09-14", "reputability": "official" },
"default_installed": false,
"safety": "Scanned at 9b794ae (2026-09-14): 764 KB, 17 skills, no executables, no hooks.json, no .claude/settings.json, no .mcp.json, no package.json lifecycle hook, no bidi or invisible Unicode, no eval, atob, new Function or createRequire, no child_process. What it drives is the pulumi CLI against whichever stack and cloud credentials your shell already holds, so pulumi up from one of these skills changes real infrastructure; its pulumi-esc skill reads and writes secrets in Pulumi ESC. It ships one symlink, CLAUDE.md -> AGENTS.md at the pack root, resolving to the real AGENTS.md beside it; see the redis entry for exactly what each copy path does with it.",
"tags": ["pulumi", "iac", "devops", "infra", "secrets", "official"]
},
{
"id": "startup-builder",
"name": "Startup Builder Skill",
"type": "skill",
"tier": "extension",
"domain": "go-to-market",
"path": ".claude/skills/ext/startup-builder",
"commit": "a738bc8f54fcc39d40647ca791a5af375025c72d",
"description": "The startup work around a Solana project, in ten skills that hand off through .claude/context/*.md files: idea-sprint (find and validate an idea, writes idea.md), build-status (build and traction status, writes build.md), positioning (ICP and alternatives), pricing (fees, floor costs, willingness to pay), pitch-deck, hackathon (track choice, submission, demo script), fundraising (data room, investor updates, token vs equity), launch (launch-day runbook with go/no-go gates), incident-comms (severity, status page, postmortem) and ecosystem-bd (Jupiter and wallet listings, partner map, grants). idea-sprint, pitch-deck and hackathon used to ship in the kit. The skills link to each other's format files, so it installs as a set.",
"triggers": ["validate a startup idea", "pitch deck or demo day", "hackathon submission", "build or traction status", "positioning or ICP", "pricing an onchain product", "fundraising data room or investor update", "launch-day runbook", "incident comms or status page", "ecosystem BD or listings"],
"source": "https://github.com/solanabr/startup-builder-skill",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add startup-builder", "env": [] },
"license": "MIT",
"maintainer": "solanabr",
"signal": { "stars": 0, "last_commit": "2026-10-06", "reputability": "first-party" },
"default_installed": false,
"safety": "Scanned at a738bc8 (2026-10-23): 97 files (90 Markdown, 3 HTML templates, LICENSE, one shell script, one Python script, one YAML workflow), 266 blobs across 80 commits. One executable, skills/pitch-review/references/anonymize.sh, the only file ever executable in its history: 138 lines of bash, run as bash anonymize.sh reading deck.html swaps.txt forbidden.txt /tmp/review-r1, that swaps real names for fake ones with perl (terms quotemeta'd, so the lists cannot inject code), strips comments, <img> and notes, prints a PDF with headless Chrome from a file:// copy and runs a leak check. The script itself writes only the empty readers' folder it is given (warning under $HOME) and a pitch-work.* folder under $TMPDIR that it never deletes, never edits its input, makes no network call and reads CHROME, CHROME_FLAGS, TMPDIR, HOME and PATH; Chrome still loads any remote asset the user's deck references, and the no-PDF hint suggests CHROME_FLAGS=--no-sandbox without anonymize.md's only-for-your-own-deck caveat. No symlinks, no gitlinks, no package.json, no hooks.json, no .mcp.json, no .claude/settings.json; no eval, atob, new Function, createRequire or child_process; no bidi or invisible Unicode, IP literals, credential names or keypair reads. The Python and YAML are upstream's own CI (validate.py: git ls-files plus a scheduled link check), inert nested in ext/. 79 lines run over 500 characters, all table rows, prose or attribution comments, none whitespace-padded; the HTML comments the range adds or edits are provenance notes, and the three HTML templates load nothing. The ten older skills write .claude/context/*.md and drafts, read public APIs (DefiLlama's api.llama.fi, verify.osec.io) and tell the agent to run read-only checks (solana program show, solana-verify get-program-hash, spl-token display). Three name unpinned npx packages, all there before this pin: hackathon and idea-sprint have the user sign in once with npx @colosseum-org/copilot-connect login for the optional colosseum-copilot pack, and ecosystem-bd's IDL step is npx @solana-program/program-metadata write idl (or anchor idl init/upgrade), signed by the upgrade authority. The launch runbook gives its eng lead the verified-build steps (a solana-verify export-pda-tx Squads proposal, verify-from-repo, which uploads the verification PDA, and remote submit-job), and the skill's launch-day mode walks those gates; set-upgrade-authority, spl-token authorize and program close appear only as warnings in slip-and-abort.md. Beyond those, the ten make no onchain or network write themselves. The two skills added at this pin reach further. pitch-review writes readers' folders outside the project (one per round, plus one for the paired comparison with a key file outside it), offers to copy a blind-reader agent into ~/.claude/agents/ (or pass it with --agents, which writes nothing), and has you open a second claude session in a readers' folder with blockReadsOutsideWorkingDirectories, outside the project, so the kit's project settings and guard hooks do not apply there. Its step 8 sends the deck to three to five outside readers through a third-party page-analytics link (Papermark or DocSend), and its paired comparison has you save four published decks from TechCrunch's SlideShare embeds by browser. market-sizing runs five web-search subagents in parallel, each given a three-line project summary and the founder-approved filters, and reads DefiLlama, the Greenhouse, Ashby and Lever job-board APIs and BLS wage data; the other hosts the range adds are citations. Content adapted from sendaifun/solana-new is credited in THIRD_PARTY_NOTICES.md without that pack's telemetry preambles; inside skills/ its only live links are to the pinned superteam-ideas.json dataset. Legal, tax and token-classification questions route to crypto-legal, which is informational only. Dated facts carry their check date; re-check one before acting on it. Pinned under ext/ rather than top-level because its twelve descriptions total 3,265 characters, about 816 tokens in every session and subagent. MIT; first-party solanabr repo, so it needs no provenance-table row.",
"tags": ["startup", "go-to-market", "pitch", "hackathon", "fundraising", "launch", "pricing", "bd"]
},
{
"id": "anthropic-claude-code-plugins",
"name": "Anthropic Claude Code Plugins",
"type": "plugin",
"domain": "dev-workflow",
"description": "Anthropic's official marketplace, 315 plugins, registered by Claude Code itself. Which ones earn their keep in a Solana project and what each costs per session: docs/configuration.md. code-review, the three LSPs, session-report, skill-creator and mcp-server-dev ship no hooks.",
"source": "https://github.com/anthropics/claude-plugins-official",
"install": { "method": "plugin-install", "command": "/plugin install code-review@claude-plugins-official", "env": [] },
"license": "Apache-2.0",
"maintainer": "anthropics",
"signal": { "stars": 37359, "last_commit": "2026-10-02", "reputability": "official" },
"default_installed": false,
"safety": "clean (Anthropic-authored; every external entry SHA-pinned). Seven first-party plugins ship hooks: hookify collides with the kit's firewall (its PreToolUse/PostToolUse hooks gate the same tool calls), and security-guidance, claude-security and code-modernization overlap /diff-review and /audit-solana at a standing context cost — docs/configuration.md has the per-plugin numbers",
"tags": ["code-review", "lsp", "session-report", "skill-creator", "official"]
},
{
"id": "get-shit-done",
"name": "Get Shit Done (GSD)",
"type": "template-repo",
"domain": "dev-workflow",
"description": "Spec-driven / meta-prompting workflow system (agents + commands).",
"source": "https://github.com/gsd-build/get-shit-done",
"install": { "method": "git-clone", "command": "git clone https://github.com/gsd-build/get-shit-done", "env": [] },
"license": "MIT",
"maintainer": "gsd-build",
"signal": { "stars": 64406, "last_commit": "2026-05-31", "reputability": "org" },
"default_installed": false,
"safety": "archived: the repository is read-only and its README points to open-gsd/gsd-core, which the kit has not reviewed; don't install this one",
"tags": ["spec-driven", "meta-prompting", "workflow", "agents", "commands"]
},
{
"id": "dev-browser",
"name": "Dev Browser",
"type": "skill",
"domain": "testing-qa",
"description": "Gives the agent a real web browser for frontend QA, testing, and scraping.",
"source": "https://github.com/SawyerHood/dev-browser",
"install": { "method": "submodule", "command": "git submodule add https://github.com/SawyerHood/dev-browser .claude/skills/ext/dev-browser", "env": [] },
"license": "MIT",
"maintainer": "SawyerHood",
"signal": { "stars": 6300, "last_commit": "2026-06-05", "reputability": "individual" },
"default_installed": false,
"safety": "clean",
"tags": ["browser", "testing", "qa", "frontend"]
},
{
"id": "ios-simulator-skill",
"name": "iOS Simulator Skill",
"type": "skill",
"domain": "testing-qa",
"description": "xcodebuild wrapper to drive the iOS Simulator; pairs with the kit's build-mobile.",
"source": "https://github.com/conorluddy/ios-simulator-skill",
"install": { "method": "submodule", "command": "git submodule add https://github.com/conorluddy/ios-simulator-skill .claude/skills/ext/ios-simulator", "env": [] },
"license": "MIT",
"maintainer": "conorluddy",
"signal": { "stars": 1100, "last_commit": "2026-06-14", "reputability": "individual" },
"default_installed": false,
"safety": "clean",
"tags": ["ios", "simulator", "xcode", "mobile", "testing"]
},
{
"id": "anydesign",
"name": "AnyDesign",
"type": "skill",
"domain": "frontend-design",
"description": "Turn an image / URL / Figma reference into design.md design tokens; closest genre-mate to animation-principles.",
"source": "https://github.com/uxKero/anydesign",
"install": { "method": "submodule", "command": "git submodule add https://github.com/uxKero/anydesign .claude/skills/ext/anydesign", "env": [] },
"license": "MIT",
"maintainer": "uxKero",
"signal": { "stars": 109, "last_commit": "2026-06-11", "reputability": "individual" },
"default_installed": false,
"safety": "clean",
"tags": ["design-tokens", "figma", "ui", "frontend"]
},
{
"id": "webgpu-claude-skill",
"name": "WebGPU Claude Skill",
"type": "skill",
"domain": "frontend-design",
"description": "WebGPU + Three.js graphics skill for the browser.",
"source": "https://github.com/dgreenheck/webgpu-claude-skill",
"install": { "method": "submodule", "command": "git submodule add https://github.com/dgreenheck/webgpu-claude-skill .claude/skills/ext/webgpu", "env": [] },
"license": "none/unspecified",
"maintainer": "dgreenheck",
"signal": { "stars": 1000, "last_commit": "2026-04-10", "reputability": "individual" },
"default_installed": false,
"safety": "clean",
"tags": ["webgpu", "three.js", "graphics", "frontend"]
},
{
"id": "swiftui-design-skill",
"name": "SwiftUI Design Skill",
"type": "skill",
"domain": "frontend-design",
"description": "Anti-AI-slop SwiftUI design rules for Apple/mobile builders.",
"source": "https://github.com/Wholiver/swiftui-design-skill",
"install": { "method": "submodule", "command": "git submodule add https://github.com/Wholiver/swiftui-design-skill .claude/skills/ext/swiftui-design", "env": [] },
"license": "MIT",
"maintainer": "Wholiver",
"signal": { "stars": 138, "last_commit": "2026-05-01", "reputability": "individual" },
"default_installed": false,
"safety": "clean",
"tags": ["swiftui", "design", "apple", "mobile", "ios"]
},
{
"id": "animation-principles",
"name": "Animation Principles",
"type": "skill",
"domain": "frontend-design",
"description": "Genre-template skill for motion / animation design principles (pure prose).",
"source": "https://github.com/dylantarre/animation-principles",
"install": { "method": "submodule", "command": "git submodule add https://github.com/dylantarre/animation-principles .claude/skills/ext/animation-principles", "env": [] },
"license": "MIT",
"maintainer": "dylantarre",
"signal": { "stars": 47, "last_commit": "2025-12-30", "reputability": "individual" },
"default_installed": false,
"safety": "clean — pure-prose, zero scripts/installers; stale (~Dec 2025), unmaintained",
"tags": ["animation", "motion", "ui", "frontend"]
},
{
"id": "brand-design-md",
"name": "Brand Design (md)",
"type": "skill",
"domain": "frontend-design",
"description": "Brand-design reference in markdown.",
"source": "https://github.com/zephyrwang6/brand-design-md",
"install": { "method": "submodule", "command": "git submodule add https://github.com/zephyrwang6/brand-design-md .claude/skills/ext/brand-design-md", "env": [] },
"license": "none/unspecified",
"maintainer": "zephyrwang6",
"signal": { "stars": 88, "last_commit": "2026-04-10", "reputability": "individual" },
"default_installed": false,
"safety": "clean; overlaps solana-new brand-design",
"tags": ["branding", "design", "frontend"]
},
{
"id": "ux-writing-skill",
"name": "UX Writing Skill",
"type": "skill",
"domain": "ux-writing",
"description": "Systematic UX microcopy: onboarding, error, empty-state and button copy with quality standards.",
"source": "https://github.com/content-designer/ux-writing-skill",
"install": { "method": "submodule", "command": "git submodule add https://github.com/content-designer/ux-writing-skill .claude/skills/ext/ux-writing", "env": [] },
"license": "MIT",
"maintainer": "content-designer",
"signal": { "stars": 112, "last_commit": "2026-05-26", "reputability": "individual" },
"default_installed": false,
"safety": "clean (one build-skill.sh builder, not runtime)",
"tags": ["ux-writing", "microcopy", "content-design", "onboarding"]
},
{
"id": "design-skills",
"name": "Design Skills (UX process)",
"type": "skill",
"domain": "ux-writing",
"description": "UX-process breadth: 10 skills incl. ux-research, design-critique, journey-mapping, a11y-audit.",
"source": "https://github.com/cuellarfr/design-skills",
"install": { "method": "submodule", "command": "git submodule add https://github.com/cuellarfr/design-skills .claude/skills/ext/design-skills", "env": [] },
"license": "MIT",
"maintainer": "cuellarfr",
"signal": { "stars": 31, "last_commit": "2026-05-04", "reputability": "individual" },
"default_installed": false,
"safety": "clean",
"tags": ["ux-research", "design-critique", "journey-mapping", "accessibility", "ux"]
},
{
"id": "scientific-agent-skills",
"name": "Scientific Agent Skills",
"type": "skill",
"domain": "data",
"description": "177 science skills + 100+ database connectors; best-in-class for data-heavy builders.",
"source": "https://github.com/K-Dense-AI/scientific-agent-skills",
"install": { "method": "submodule", "command": "git submodule add https://github.com/K-Dense-AI/scientific-agent-skills .claude/skills/ext/scientific-agent-skills", "env": [] },
"license": "MIT",
"maintainer": "K-Dense-AI",
"signal": { "stars": 47355, "last_commit": "2026-10-01", "reputability": "org" },
"default_installed": false,
"safety": "clean",
"tags": ["science", "data", "database-connectors", "research"]
},
{
"id": "claude-d3js-skill",
"name": "Claude D3.js Skill",
"type": "skill",
"domain": "data",
"description": "D3 data-visualization skill for dashboards.",
"source": "https://github.com/chrisvoncsefalvay/claude-d3js-skill",
"install": { "method": "submodule", "command": "git submodule add https://github.com/chrisvoncsefalvay/claude-d3js-skill .claude/skills/ext/d3js", "env": [] },
"license": "none/unspecified",
"maintainer": "chrisvoncsefalvay",
"signal": { "stars": 192, "last_commit": "2025-10-18", "reputability": "individual" },
"default_installed": false,
"safety": "clean; stale (~Oct 2025)",
"tags": ["d3", "data-viz", "dashboards", "data"]
},
{
"id": "claude-skills-marketplace",
"name": "Claude Skills Marketplace (mhattingpete)",
"type": "plugin",
"domain": "dev-workflow",
"description": "Plugin marketplace: git + test + code-review pack.",
"source": "https://github.com/mhattingpete/claude-skills-marketplace",
"install": { "method": "plugin-marketplace", "command": "/plugin marketplace add mhattingpete/claude-skills-marketplace", "env": [] },
"license": "Apache-2.0",
"maintainer": "mhattingpete",
"signal": { "stars": 607, "last_commit": "2026-03-06", "reputability": "individual" },
"default_installed": false,
"safety": "clean; borderline stale (~Mar 2026)",
"tags": ["git", "testing", "code-review", "marketplace", "workflow"]
},
{
"id": "playwright-skill",
"name": "Playwright Skill",
"type": "skill",
"tier": "extension",
"domain": "testing-qa",
"path": ".claude/skills/ext/playwright-skill",
"commit": "dd47a6a023e249eb1b36e9e943eab89d0900865d",
"description": "Browser automation and end-to-end webapp tests driven by Playwright: a test executor plus the selector, waiting and assertion patterns that keep a suite from going flaky. Layers above the core anthropic-skills webapp-testing skill rather than replacing it — reach for this one when a flow needs a persistent browser session, multiple contexts or a real CI suite rather than a one-off check.",
"triggers": ["Playwright", "end-to-end test", "browser automation", "e2e suite", "flaky selector"],
"source": "https://github.com/lackeyjb/playwright-skill",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add playwright-skill", "env": [] },
"license": "MIT",
"maintainer": "lackeyjb",
"signal": { "stars": 3170, "last_commit": "2026-08-14", "reputability": "individual" },
"default_installed": false,
"safety": "Scanned at dd47a6a (2026-08-14): 136 KB, one skill folder, no symlinks, no hooks.json, no .claude/settings.json, no .mcp.json, no package.json install lifecycle hook, no bidi or invisible Unicode, no lines over 500 characters, no eval, atob, new Function or createRequire. One executable, skills/playwright-skill/run.js, the test executor the skill drives; it and tests/executor.test.js use child_process to launch a browser and run the test command you give it, which is the pack's whole purpose. It drives a real browser against whatever URL you point it at, so a test pointed at production acts on production. Individually authored, so it carries a row in the provenance table.",
"tags": ["playwright", "browser", "testing", "qa", "e2e"]
},
{
"id": "vercel-plugin",
"name": "Vercel Plugin",
"type": "plugin",
"domain": "frontend-design",
"description": "Full Vercel plugin: 28 skills + 3 agents + official read-only Vercel MCP (Next.js/React/AI-SDK frontend bundle).",
"source": "https://github.com/vercel/vercel-plugin",
"install": { "method": "plugin-marketplace", "command": "/plugin marketplace add vercel/vercel-plugin", "env": ["VERCEL_PLUGIN_TELEMETRY=off (opt-out of default telemetry)"] },
"license": "Apache-2.0",
"maintainer": "vercel",
"signal": { "stars": 295, "last_commit": "2026-10-02", "reputability": "org" },
"default_installed": false,
"safety": "caution: telemetry ON by default (anonymized DAU pings, no content); opt-out VERCEL_PLUGIN_TELEMETRY=off",
"tags": ["vercel", "nextjs", "react", "ai-sdk", "frontend", "deployment"]
},
{
"id": "meteora-invent",
"name": "Meteora Agent Skill",
"type": "skill",
"tier": "extension",
"domain": "solana-defi",
"path": ".claude/skills/ext/meteora-invent",
"commit": "99d1441e8e39a11752a13c16c12867dba0448c9d",
"description": "Meteora's official skill (skills/meteora in the meteora-invent monorepo): DBC launches, DAMM v1/v2 and DLMM pools, alpha and presale vaults, locks and vesting, M3M3 staking and zaps, with version-pinned SDK references. Deeper than the sendai meteora folder, which it supersedes for Meteora work.",
"triggers": ["Meteora", "DLMM", "DAMM", "dynamic bonding curve", "alpha vault", "M3M3 staking"],
"source": "https://github.com/MeteoraAg/meteora-invent",
"install": { "method": "kit", "command": "bash .claude/bin/skills.sh add meteora-invent", "env": [] },
"license": "MIT",
"maintainer": "MeteoraAg",
"signal": { "stars": 34, "last_commit": "2026-09-20", "reputability": "official" },
"default_installed": false,
"safety": "Scanned at dd77ef3 (2026-09-20): no symlinks, no hooks.json, no .claude/settings.json, no .mcp.json, no bidi or invisible Unicode. The skill itself (skills/meteora, about 200 KB) is SKILL.md plus references with no scripts; its on-chain writes need a funded keypair and it asks for owner confirmation first. The pin carries the whole monorepo, which is the cost of pinning it: 19 MB checked out, including six prebuilt .so program artifacts under studio/src/tests/artifacts (cp_amm, dynamic_bonding_curve, lb_clmm, locker, alpha_vault, transfer_hook_counter) that exist for the repo's own LiteSVM tests and that nothing in the skill invokes. Its root package.json declares prepare: husky install, a lifecycle hook that would run only if someone ran npm install inside the pack directory; no installer the kit ships does that. Previously an add-on for the size reason — the blobless folder-fetch route that would take skills/meteora alone is blocked by skills.sh's requirement of an Apache-2.0 LICENSE.txt inside each fetched folder, which this repo does not carry per-folder.",
"tags": ["meteora", "dlmm", "defi", "sdk", "solana"]
},
{
"id": "solana-airdrop",
"name": "Solana Airdrop (Merkle distribution)",
"type": "skill",
"domain": "solana-defi",
"description": "Merkle-tree airdrops: recipient ingestion, root and per-recipient proofs, claim-instruction building for the Solana Foundation template and Metaplex gumdrop, anti-double-claim PDA derivation, anti-sybil eligibility patterns. A zero-dependency pure-Python Keccak-256 engine plus a TypeScript claim builder.",
"triggers": ["airdrop", "Merkle distribution", "claim contract", "recipient proofs", "whitelist eligibility"],
"source": "https://github.com/yusizer/solana-airdrop-skill",
"install": { "method": "git-clone", "command": "git clone --depth 1 https://github.com/yusizer/solana-airdrop-skill && cd solana-airdrop-skill && bash install.sh .", "env": [] },
"license": "MIT",
"maintainer": "yusizer",
"signal": { "stars": 0, "last_commit": "2026-06-27", "reputability": "individual" },
"default_installed": false,
"safety": "no malware: scanned at 1b708843cd8104e3dd4433a0e85a6b443ed4c711 — no eval, atob or createRequire, no network egress, no keypair reads, stdlib-only Python. Its Keccak-256 is correct (cross-checked against SHA3 over 305 input lengths). Correctness caveats, though, and they matter because airdrop roots are irreversible: verify_all() is self-referential and cannot fail, so it does not catch a wrong leaf encoding, and its eval grades itself against its own root. De-dup keys on (pubkey, amount) while the Foundation template's claim PDA is per-claimant, so a repeated pubkey strands funds. The claim builder asks for a Keypair where a PublicKey would do, and its install.sh overwrites a project's agents, commands and rules without asking. See upstream issue 1; treat its gate as a self-consistency check and verify a root against the target program's own tooling before publishing",
"tags": ["airdrop", "merkle", "distribution", "claim", "keccak"]
},
{
"id": "buildwithclaude",
"name": "Build With Claude",
"type": "aggregator",
"domain": "dev-workflow",
"description": "Scout source: live searchable hub (buildwithclaude.com); best for plugins/marketplaces. Not installable.",
"source": "https://github.com/davepoon/buildwithclaude",
"install": null,
"license": "MIT",
"maintainer": "davepoon",
"signal": { "stars": 3576, "last_commit": "2026-10-02", "reputability": "individual" },
"default_installed": false,
"safety": "clean — discovery source only, nothing executed",
"tags": ["aggregator", "scout", "plugins", "marketplaces", "discovery"]
},
{
"id": "voltagent-awesome-claude-code-subagents",
"name": "VoltAgent Awesome Claude Code Subagents",
"type": "aggregator",
"domain": "dev-workflow",
"description": "Scout source: 100+ subagents by domain (agents, not skills). Not installable.",
"source": "https://github.com/VoltAgent/awesome-claude-code-subagents",
"install": null,
"license": "MIT",
"maintainer": "VoltAgent",
"signal": { "stars": 25457, "last_commit": "2026-09-21", "reputability": "org" },
"default_installed": false,
"safety": "clean — discovery source only, nothing executed",
"tags": ["aggregator", "scout", "subagents", "discovery"]
},
{
"id": "voltagent-awesome-agent-skills",
"name": "VoltAgent Awesome Agent Skills",
"type": "aggregator",
"domain": "dev-workflow",
"description": "Scout source: skills index whose \"Official Skills by\" section lists vendor-published packs. Not installable.",
"source": "https://github.com/VoltAgent/awesome-agent-skills",
"install": null,
"license": "MIT",
"maintainer": "VoltAgent",
"signal": { "stars": 35191, "last_commit": "2026-10-02", "reputability": "org" },
"default_installed": false,
"safety": "clean \u2014 discovery source only, nothing executed. Carries a licence file, unlike several larger skill indexes",
"tags": ["aggregator", "scout", "skills", "vendor-packs", "discovery"]
},
{
"id": "github-awesome-copilot",
"name": "GitHub Awesome Copilot",
"type": "aggregator",
"domain": "dev-workflow",
"description": "Scout source: GitHub's own index of Copilot and agent customisations \u2014 instructions, prompts and chat modes. Not installable.",
"source": "https://github.com/github/awesome-copilot",
"install": null,
"license": "MIT",
"maintainer": "github",
"signal": { "stars": 39687, "last_commit": "2026-10-01", "reputability": "official" },
"default_installed": false,
"safety": "clean \u2014 discovery source only, nothing executed",
"tags": ["aggregator", "scout", "prompts", "instructions", "discovery"]
},
{
"id": "anthropics-claude-plugins-community",
"name": "Anthropic Claude Plugins Community Index",
"type": "aggregator",
"domain": "dev-workflow",
"description": "Scout source: 2,201-plugin community index; discovery source, NOT first-party-vetted. Not installable.",
"source": "https://github.com/anthropics/claude-plugins-community",
"install": null,
"license": "Apache-2.0",
"maintainer": "anthropics",
"signal": { "stars": 4451, "last_commit": "2026-10-01", "reputability": "official" },
"default_installed": false,
"safety": "clean — community index, not first-party-vetted; verify each entry before installing",
"tags": ["aggregator", "scout", "plugins", "community", "discovery"]
},
{
"id": "coderabbit",
"name": "CodeRabbit Skills",
"type": "skill",
"domain": "dev-workflow",
"description": "CodeRabbit's own skills for driving its AI code review: configuring review scope and path filters, reading and acting on review comments, and the CLI. Catalogued rather than pinned because the kit's review lane is already covered end to end by /diff-review, /audit-solana and auditor-skill, and this one only earns its keep on a repo that has CodeRabbit installed on its pull requests.",
"source": "https://github.com/coderabbitai/skills",
"install": { "method": "submodule", "command": "git submodule add https://github.com/coderabbitai/skills .claude/skills/ext/coderabbit", "env": [] },
"license": "MIT",
"maintainer": "coderabbitai",
"signal": { "stars": 187, "last_commit": "2026-10-01", "reputability": "official" },
"default_installed": false,
"safety": "Not pinned and not scanned at a commit, so read it before installing. It is the vendor's own pack for a service that reads your repository and posts on your pull requests; what it can see is whatever you have already granted the CodeRabbit app.",
"tags": ["code-review", "ci", "pull-request", "dev-workflow"]
},
{
"id": "dagster",
"name": "Dagster Skills",
"type": "skill",
"domain": "data-analytics",
"description": "Dagster's own skills for building and operating asset-based data pipelines: asset and job definitions, partitions and backfills, schedules and sensors, and debugging a failed run. Catalogued rather than pinned because an orchestrator is a heavier commitment than most Solana indexers need — a webhook consumer plus a scheduled job usually is the pipeline. Reach for it when the pipeline has grown real dependency graphs and backfills.",
"source": "https://github.com/dagster-io/skills",
"install": { "method": "submodule", "command": "git submodule add https://github.com/dagster-io/skills .claude/skills/ext/dagster", "env": [] },
"license": "Apache-2.0",
"maintainer": "dagster-io",
"signal": { "stars": 210, "last_commit": "2026-10-02", "reputability": "official" },
"default_installed": false,
"safety": "Not pinned and not scanned at a commit, so read it before installing. Its default branch is master rather than main, which matters if you pin it later.",
"tags": ["dagster", "orchestration", "pipeline", "data", "etl"]
},
{
"id": "motherduck",
"name": "MotherDuck Agent Skills",
"type": "skill",
"domain": "data-analytics",
"description": "MotherDuck's own skills for its hosted DuckDB: connecting, moving a local DuckDB database into the cloud, shared databases and query patterns. Catalogued rather than pinned because the pinned duckdb extension already covers the engine and local files, and this adds only the hosted service on top — useful once an indexer's exports outgrow one machine.",
"source": "https://github.com/motherduckdb/agent-skills",
"install": { "method": "submodule", "command": "git submodule add https://github.com/motherduckdb/agent-skills .claude/skills/ext/motherduck", "env": ["motherduck_token"] },
"license": "none",
"maintainer": "motherduckdb",
"signal": { "stars": 62, "last_commit": "2026-09-06", "reputability": "official" },
"default_installed": false,
"safety": "Not pinned and not scanned at a commit, so read it before installing. The repository carries no licence file. It talks to a hosted service authenticated by a `motherduck_token`, so the data you query leaves your machine; the pinned `duckdb` extension stays entirely local.",
"tags": ["motherduck", "duckdb", "hosted", "sql", "data"]
},
{
"id": "stripe-ai",
"name": "Stripe AI Skills",
"type": "skill",
"domain": "payments",
"description": "Stripe's own pack (providers/agent-plugins/plugin/skills): ten skills over Stripe Pay, Connect onboarding and verification requirements, Stripe Apps, the Smart Contract-free fiat rails, API best practices, docs search, project scaffolding, Metronome usage billing and an upgrade path. The card-and-bank-rails counterpart to the pinned circle extension, which covers USDC and stablecoins; reach for this one when the money has to leave or enter the traditional system.",
"source": "https://github.com/stripe/ai",
"install": { "method": "git-clone", "command": "git clone --depth 1 https://github.com/stripe/ai stripe-ai && cp -R stripe-ai/providers/agent-plugins/plugin/skills/stripe-pay .claude/skills/stripe-pay", "env": ["STRIPE_SECRET_KEY"] },
"license": "MIT",
"maintainer": "stripe",
"signal": { "stars": 1855, "last_commit": "2026-10-05", "reputability": "official" },
"default_installed": false,
"safety": "Not pinned and not scanned at a commit, so read it before installing; it is an add-on for two mechanical reasons rather than anything about the content, and both are worth knowing before anyone retries a pin. First, a submodule pin is unavailable: the repo ships four self-referential symlinks (`llm/ai-sdk/LICENSE`, `llm/token-meter/LICENSE`, `tools/modelcontextprotocol/LICENSE`, `tools/typescript/LICENSE`, each pointing at `LICENSE` in its own directory), so they resolve to nothing and `cp` aborts on them with `Too many levels of symbolic links`. The copy sites that would hit them sit in the byte-frozen region of `.claude/bin/update.sh`, which an existing install runs from its own old copy, so a pin would break `/update` for every already-installed project with no way to fix it from upstream. Second, the kit's folder-fetch route does avoid those links — they are outside `providers/agent-plugins/plugin/skills/`, and that route checks out only the named folders — but `check_skill()` in `.claude/bin/skills.sh` requires an Apache-2.0 `LICENSE.txt` inside each fetched folder and refuses a folder that has none; stripe's skill folders contain only `SKILL.md`. The install command above sidesteps both by copying one skill folder out of a shallow clone, which never walks `llm/` or `tools/`. On the content itself: it moves real money through your Stripe account once a secret key is set, and `providers/` carries runnable TypeScript and MCP tooling beyond the skill text, so read the folder you copy rather than the repo wholesale.",
"tags": ["stripe", "payments", "fiat", "billing", "connect", "official"]
}
]
}
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |